Posts

Coffee Shop Location Predictor

As part of this article, we will explore the main steps involved in predicting the best location for a coffee shop in Vancouver. We will also take into consideration that the coffee shop is near a transit station, and has no Starbucks near it. Well, while at it, let us also add an extra feature where we make sure the crime in the area is lower.

Introduction

In this article, we will highlight the main steps involved to predict a location for a coffee shop in Vancouver. We also want to make sure that the coffee shop is near a transit station, and has no Starbucks near it. As an added feature, we will make sure that the crime concentration in the area is low, and the entire program should be implemented in Python. So let’s walk through the steps.

Steps Required

  • Get crime history for the last two years
  • Get locations of all transit stations and Starbucks in Vancouver
  • Check all the transit stations that do not have any Starbucks near them
  • Get all the data regarding crimes near the filtered transit stations
  • Create a grid of all possible coordinates around the transit station
  • Check crime around each created coordinate and display the top 5 locations.

Gathering Data

This covers the first two steps required to get data from the internet, both manually and automatically.

Getting all Crime History

We can get crime history for the past 14 years in Vancouver from here. This data is in raw crime.csv format, so we have to process it and filter out useless data. We then write this processed information on the crime_processed.csv file.

Note: There are 530,653 records of crime in this file

In this program, we will just use the type and coordinate of the crime. There are many crime types, but we have classified them into three major categories namely;

Theft (red), Break and Enter (orange) and Mischief (green)

These all crimes can be plotted on Graph as displayed below.

This may seem very congested and full, so let’s see a closeup image for future references.

Getting Locations of all Rapid Transit Stations

We can get the coordinates of all Transit Stations in Vancouver from here. This dataset has all coordinates of rapid transit stations in three transit lines in Vancouver. There are a total of 23 of them in Vancouver, we can then use it for further processing.

Getting Locations of all Starbucks

The Starbucks data is present here, we can scrape it easily and get the locations of all the Starbucks in Vancouver. We just need the Starbucks that is near transit stations, so we’ll filter out the rest. There are a total 24 Starbucks in Vancouver, and 10 of them are near Transit Stations.

Note: Other than the coordinates of Transit Stations and Starbucks, we also need coordinates and type of the crime.

Transit Stations with no Starbucks

As we have all the data required, now moving to the next step. We need to get to the transit Station locations that have no Starbucks near them. For that we can create an area of particular radius around each Transit Station. Then check all Starbucks locations with respect to them, whether they are within that area or not.

If none of the Starbucks are within that particular Transit Station’s area, we can append it to a list. At the end, we have a list of all Transit locations with no Starbucks near them. There are a total of 6 Transit Stations with no Starbucks near them.

Crime near Transit Stations

Now lets filter out all crime records and get just what we are interested in, which means the crime near Transit stations. For that we will plot an area of specific radius around each of them to see the crimes. These are more than 110,000 crime records.

Crime near located Transit Stations

Now that we have all the Transit Stations that don’t have any Starbucks near them and also the crime near all Transit Stations. So, let’s use this information and get crime near the located Transit Stations. These are about 44,000 crime records.

This may seem correct at first glance, but the points are overlapping due to abundance, so we can create different lists of crimes based on their types.

Theft

Break and Enter

Mischief

Generating all possible coordinates

Now finally, we have all the prerequisites and let’s get to the main task at hand, predicting the best coordinate for the coffee shop.

There may be many approaches to solve this problem, but the one I used in this program is that I will create a grid of all possible locations (coordinates) in the area of 1 km radius around each located transit station.

Initially I generated 1 coordinate for every m, this resulted in 1000,000 coordinates in every km. This is a huge number, and for the 6 located Transit stations, it becomes 6 Million. It may not seem much at first glance because computers can handle such data in a few seconds.

But for location prediction we need to compare each coordinate with crime coordinates. As the algorithm has to check for ~7,000 Thefts, ~19,000 Break ins, and ~17,000 Mischiefs around each generated coordinate. Computing this would want the program to process an estimate of 432.4 Billion times. This sort of execution takes many hours on normal computers (sometimes days).

The solution to this is to create a coordinate for each 10 m area, this results about 10,000 coordinate per km. For the above mentioned number of crimes, the estimated processes will be several Billions. That would significantly reduce the time, but is still not less.

To control this, we can remove the duplicate values in crime coordinates and those which are too close to each other ~1m. Doing so, we are left with just 816 Thefts, 2,654 Break ins, and 8,234 Mischiefs around each generated coordinate.
The precision will not be affected much but the time and computational resources required will be reduced a lot.

 

Checking Crime near Generated coordinates

Now that we have all the locations, we will start some processing on it and check each coordinate against some constraints. That are respectively;

  1. Filter out Coordinates having Theft near 1 km
    We get 122,000 coordinates with no Thefts (Below merged 1000 to 1)
  2. Filter out Coordinates having Break Ins near 200m
    We get 8000 coordinates with no Thefts (Below merged 1000 to 1)
  3. Filter out Coordinates having Mischief near 200m
    We get 6000 coordinates with no Thefts (Below merged 1000 to 1)
    Now that we have 6 Coordinates of best locations that have passed through all the constraints, we will order them.To order them, we will check their distance from the nearest transit location. The nearest will be on top of the list as the best possible location, then the second and so on. The generated List is;

    1. -123.0419406741792, 49.24824259252004
    2. -123.05887151659479, 49.24327221040713
    3. -123.05287151659476, 49.24327221040713
    4. -123.04994067417924, 49.239242592520064
    5. -123.0419406741792, 49.239242592520064
    6. -123.0409406741792, 49.239242592520064

How can MindTrades help?

MindTrades Consulting Services, a leading marketing agency provides in-depth analysis and insights for the global IT sector including leading data integration brands such as Diyotta. From Cloud Migration, Big Data, Digital Transformation, Agile Deliver, Cyber Security, to Analytics- Mind trades provides published breakthrough ideas, and prompt content delivery. For more information, refer to mindtrades.com.

Code

https://github.com/Mindtrades-Consulting/Coffee-Shop-Location-Predictor

 

Rethinking linear algebra part two: ellipsoids in data science

*This is the fourth article of my article series “Illustrative introductions on dimension reduction.”

1 Our expedition of eigenvectors still continues

This article is still going to be about eigenvectors and PCA, and this article still will not cover LDA (linear discriminant analysis). Hereby I would like you to have more organic links of the data science ideas with eigenvectors.

In the second article, we have covered the following points:

  • You can visualize linear transformations with matrices by calculating displacement vectors, and they usually look like vectors swirling.
  • Diagonalization is finding a direction in which the displacement vectors do not swirl, and that is equal to finding new axis/basis where you can describe its linear transformations more straightforwardly. But we have to consider diagonalizability of the matrices.
  • In linear dimension reduction such as PCA or LDA, we mainly use types of matrices called positive definite or positive semidefinite matrices.

In the last article we have seen the following points:

  • PCA is an algorithm of calculating orthogonal axes along which data “swell” the most.
  • PCA is equivalent to calculating a new orthonormal basis for the data where the covariance between components is zero.
  • You can reduced the dimension of the data in the new coordinate system by ignoring the axes corresponding to small eigenvalues.
  • Covariance matrices enable linear transformation of rotation and expansion and contraction of vectors.

I emphasized that the axes are more important than the surface of the high dimensional ellipsoids, but in this article let’s focus more on the surface of ellipsoids, or I would rather say general quadratic curves. After also seeing how to draw ellipsoids on data, you would see the following points about PCA or eigenvectors.

  • Covariance matrices are real symmetric matrices, and also they are positive semidefinite. That means you can always diagonalize covariance matrices, and their eigenvalues are all equal or greater than 0.
  • PCA is equivalent to finding axes of quadratic curves in which gradients are biggest. The values of quadratic curves increases the most in those directions, and that means the directions describe great deal of information of data distribution.
  • Intuitively dimension reduction by PCA is equal to fitting a high dimensional ellipsoid on data and cutting off the axes corresponding to small eigenvalues.

Even if you already understand PCA to some extent, I hope this article provides you with deeper insight into PCA, and at least after reading this article, I think you would be more or less able to visually control eigenvectors and ellipsoids with the Numpy and Maplotlib libraries.

*Let me first introduce some mathematical facts and how I denote them throughout this article in advance. If you are allergic to mathematics, take it easy or please go back to my former articles.

  • Any quadratic curves can be denoted as \boldsymbol{x}^T A\boldsymbol{x} + 2\boldsymbol{b}^T\boldsymbol{x} + s = 0, where \boldsymbol{x}\in \mathbb{R}^D , A \in \mathbb{R}^{D\times D} \boldsymbol{b}\in \mathbb{R}^D s\in \mathbb{R}.
  • When I want to clarify dimensions of variables of quadratic curves, I denote parameters as A_D, b_D.
  • If a matrix A is a real symmetric matrix, there exist a rotation matrix U such that U^T A U = \Lambda, where \Lambda = diag(\lambda_1, \dots, \lambda_D) and U = (\boldsymbol{u}_1, \dots , \boldsymbol{u}_D). \boldsymbol{u}_1, \dots , \boldsymbol{u}_D are eigenvectors corresponding to \lambda_1, \dots, \lambda_D respectively.
  • PCA corresponds to a case of diagonalizing A where A is a covariance matrix of certain data. When I want to clarify that A is a covariance matrix, I denote it as A=\Sigma.
  • Importantly covariance matrices \Sigma are positive semidefinite and real symmetric, which means you can always diagonalize \Sigma and any of their engenvalues cannot be lower than 0.

*In the last article, I denoted the covariance of data as S, based on Pattern Recognition and Machine Learning by C. M. Bishop.

*Sooner or later you are going to see that I am explaining basically the same ideas from different points of view, using the topic of PCA. However I believe they are all important when you learn linear algebra for data science of machine learning. Even you have not learnt linear algebra or if you have to teach linear algebra, I recommend you to first take a review on the idea of diagonalization, like the second article. And you should be conscious that, in the context of machine learning or data science, only a very limited type of matrices are important, which I have been explaining throughout this article.

2 Rotation or projection?

In this section I am going to talk about basic stuff found in most textbooks on linear algebra. In the last article, I mentioned that if A is a real symmetric matrix, you can diagonalize A with a rotation matrix U = (\boldsymbol{u}_1 \: \cdots \: \boldsymbol{u}_D), such that U^{-1}AU = U^{T}AU =\Lambda, where \Lambda = diag(\lambda_{1}, \dots , \lambda_{D}). I also explained that PCA is a case where A=\Sigma, that is, A is the covariance matrix of certain data. \Sigma is known to be positive semidefinite and real symmetric. Thus you can always diagonalize \Sigma and any of their engenvalues cannot be lower than 0.

I think we first need to clarify the difference of rotation and projection. In order to visualize the ideas, let’s consider a case of D=3. Assume that you have got an orthonormal rotation matrix U = (\boldsymbol{u}_1 \: \boldsymbol{u}_2 \: \boldsymbol{u}_3) which diagonalizes A. In the last article I said diagonalization is equivalent to finding new orthogonal axes formed by eigenvectors, and in the case of this section you got new orthonoramal basis (\boldsymbol{u}_1, \boldsymbol{u}_2, \boldsymbol{u}_3) which are in red in the figure below. Projecting a point \boldsymbol{x} = (x, y, z) on the new orthonormal basis is simple: you just have to multiply \boldsymbol{x} with U^T. Let U^T \boldsymbol{x} be (x', y', z')^T, and then \left( \begin{array}{c} x' \\ y' \\ z' \end{array} \right) = U^T\boldsymbol{x} = \left( \begin{array}{c} \boldsymbol{u}_1^{T}\boldsymbol{x} \\ \boldsymbol{u}_2^{T}\boldsymbol{x} \\ \boldsymbol{u}_3^{T}\boldsymbol{x} \end{array} \right). You can see x', y', z' are \boldsymbol{x} projected on \boldsymbol{u}_1, \boldsymbol{u}_2, \boldsymbol{u}_3 respectively, and the left side of the figure below shows the idea. When you replace the orginal orthonormal basis (\boldsymbol{e}_1, \boldsymbol{e}_2, \boldsymbol{e}_3) with (\boldsymbol{u}_1, \boldsymbol{u}_2, \boldsymbol{u}_3) as in the right side of the figure below, you can comprehend the projection as a rotation from (x, y, z) to (x', y', z') by a rotation matrix U^T.

Next, let’s see what rotation is. In case of rotation, you should imagine that you rotate the point \boldsymbol{x} in the same coordinate system, rather than projecting to other coordinate system. You can rotate \boldsymbol{x} by multiplying it with U. This rotation looks like the figure below.

In the initial position, the edges of the cube are aligned with the three orthogonal black axes (\boldsymbol{e}_1,  \boldsymbol{e}_2 , \boldsymbol{e}_3), with one corner of the cube located at the origin point of those axes. The purple dot denotes the corner of the cube directly opposite the origin corner. The cube is rotated in three dimensions, with the origin corner staying fixed in place. After the rotation with a pivot at the origin, the edges of the cube are now aligned with a new set of orthogonal axes (\boldsymbol{u}_1,  \boldsymbol{u}_2 , \boldsymbol{u}_3), shown in red. You might understand that more clearly with an equation: U\boldsymbol{x} = (\boldsymbol{u}_1 \: \boldsymbol{u}_2 \: \boldsymbol{u}_3) \left( \begin{array}{c} x \\ y \\ z \end{array} \right) = x\boldsymbol{u}_1 + y\boldsymbol{u}_2 + z\boldsymbol{u}_3. In short this rotation means you keep relative position of \boldsymbol{x}, I mean its coordinates (x, y, z), in the new orthonormal basis. In this article, let me call this a “cube rotation.”

The discussion above can be generalized to spaces with dimensions higher than 3. When U \in \mathbb{R}^{D \times D} is an orthonormal matrix and a vector \boldsymbol{x} \in \mathbb{R}^D, you can project \boldsymbol{x} to \boldsymbol{x}' = U^T \boldsymbol{x}or rotate it to \boldsymbol{x}'' = U \boldsymbol{x}, where \boldsymbol{x}' = (x_{1}', \dots, x_{D}')^T and \boldsymbol{x}'' = (x_{1}'', \dots, x_{D}'')^T. In other words \boldsymbol{x} = U \boldsymbol{x}', which means you can rotate back \boldsymbol{x}' to the original point \boldsymbol{x} with the rotation matrix U.

I think you at least saw that rotation and projection are basically the same, and that is only a matter of how you look at the coordinate systems. But I would say the idea of projection is more important through out this article.

Let’s consider a function f(\boldsymbol{x}; A) = \boldsymbol{x}^T A \boldsymbol{x} = (\boldsymbol{x}, A \boldsymbol{x}), where A\in \mathbb{R}^{D\times D} is a real symmetric matrix. The distribution of f(\boldsymbol{x}; A) is quadratic curves whose center point covers the origin, and it is known that you can express this distribution in a much simpler way using eigenvectors. When you project this function on eigenvectors of A, that is when you substitute U \boldsymbol{x}' for \boldsymbol{x}, you get f = (\boldsymbol{x}, A \boldsymbol{x}) =(U \boldsymbol{x}', AU \boldsymbol{x}') = (\boldsymbol{x}')^T U^TAU \boldsymbol{x}' = (\boldsymbol{x}')^T \Lambda \boldsymbol{x}' = \lambda_1 ({x'}_1)^2 + \cdots + \lambda_D ({x'}_D)^2. You can always diagonalize real symmetric matrices, so the formula implies that the shapes of quadratic curves largely depend on eigenvectors. We are going to see this in detail in the next section.

*(\boldsymbol{x}, \boldsymbol{y}) denotes an inner product of \boldsymbol{x} and \boldsymbol{y}.

*We are going to see details of the shapes of quadratic “curves” or “functions” in the next section.

To be exact, you cannot naively multiply U or U^T for rotation. Let’s take a part of data I showed in the last article as an example. In the figure below, I projected data on the basis (\boldsymbol{u}_1,  \boldsymbol{u}_2 , \boldsymbol{u}_3).

You might have noticed that you cannot do a “cube rotation” in this case. If you make the coordinate system (\boldsymbol{u}_1, \boldsymbol{u}_2, \boldsymbol{u}_3) with your left hand, like you might have done in science classes in school to learn Fleming’s rule, you would soon realize that the coordinate systems in the figure above do not match. You need to flip the direction of one axis to match them.

Mathematically, you have to consider the determinant of the rotation matrix U. You can do a “cube rotation” when det(U)=1, and in the case above det(U) was -1, and you needed to flip one axis to make the determinant 1. In the example in the figure below, you can match the basis. This also can be generalized to higher dimensions, but that is also beyond the scope of this article series. If you are really interested, you should prepare some coffee and snacks and textbooks on linear algebra, and some weekends.

When you want to make general ellipsoids in a 3d space on Matplotlib, you can take advantage of rotation matrices. You first make a simple ellipsoid symmetric about xyz axis using polar coordinates, and you can rotate the whole ellipsoid with rotation matrices. I made some simple modules for drawing ellipsoid. If you put in a rotation matrix which diagonalize the covariance matrix of data and a list of three radiuses \sqrt{\lambda_1}, \sqrt{\lambda_2}, \sqrt{\lambda_3}, you can rotate the original ellipsoid so that it fits the data well.

3 Types of quadratic curves.

*This article might look like a mathematical writing, but I would say this is more about computer science. Please tolerate some inaccuracy in terms of mathematics. I gave priority to visualizing necessary mathematical ideas in my article series. If you are not sure about details, please let me know.

In linear dimension reduction, or at least in this article series you mainly have to consider ellipsoids. However ellipsoids are just one type of quadratic curves. In the last article, I mentioned that when the center of a D dimensional ellipsoid is the origin point of a normal coordinate system, the formula of the surface of the ellipsoid is as follows: (\boldsymbol{x}, A\boldsymbol{x})=1, where A satisfies certain conditions. To be concrete, when (\boldsymbol{x}, A\boldsymbol{x})=1 is the surface of a ellipsoid, A has to be diagonalizable and positive definite.

*Real symmetric matrices are diagonalizable, and positive definite matrices have only positive eigenvalues. Covariance matrices \Sigma, whose displacement vectors I visualized in the last two articles, are known to be symmetric real matrices and positive semi-defintie. However, the surface of an ellipsoid which fit the data is \boldsymbol{x}^T \Sigma ^{-1} \boldsymbol{x} = const., not \boldsymbol{x}^T \Sigma \boldsymbol{x} = const..

*You have to keep it in mind that \boldsymbol{x} are all deviations.

*You do not have to think too much about what the “semi” of the term “positive semi-definite” means fow now.

As you could imagine, this is just one simple case of richer variety of graphs. Let’s consider a 3-dimensional space. Any quadratic curves in this space can be denoted as ax^2 + by^2 + cz^2 + dxy + eyz + fxz + px + qy + rz + s = 0, where at least one of a, b, c, d, e, f, p, q, r, s is not 0.  Let \boldsymbol{x} be (x, y, z)^T, then the quadratic curves can be simply denoted with a 3\times 3 matrix A and a 3-dimensional vector \boldsymbol{b} as follows: \boldsymbol{x}^T A\boldsymbol{x} + 2\boldsymbol{b}^T\boldsymbol{x} + s = 0, where A = \left( \begin{array}{ccc} a & \frac{d}{2} & \frac{f}{2} \\ \frac{d}{2} & b & \frac{e}{2} \\ \frac{f}{2} & \frac{e}{2} & c \end{array} \right), \boldsymbol{b} = \left( \begin{array}{c} \frac{p}{2} \\ \frac{q}{2} \\ \frac{r}{2} \end{array} \right). General quadratic curves are roughly classified into the 9 types below.

You can shift these quadratic curves so that their center points come to the origin, without rotation, and the resulting curves are as follows. The curves can be all denoted as \boldsymbol{x}^T A\boldsymbol{x}.

As you can see, A is a real symmetric matrix. As I have mentioned repeatedly, when all the elements of a D \times D symmetric matrix A are real values and its eigen values are \lambda_{i} (i=1, \dots , D), there exist orthogonal/orthonormal matrices U such that U^{-1}AU = \Lambda, where \Lambda = diag(\lambda_{1}, \dots , \lambda_{D}). Hence, you can diagonalize the A = \left( \begin{array}{ccc} a & \frac{d}{2} & \frac{f}{2} \\ \frac{d}{2} & b & \frac{e}{2} \\ \frac{f}{2} & \frac{e}{2} & c \end{array} \right) with an orthogonal matrix U. Let U be an orthogonal matrix such that U^T A U = \left( \begin{array}{ccc} \alpha  & 0 & 0 \\ 0 & \beta & 0 \\ 0 & 0 & \gamma \end{array} \right) =\left( \begin{array}{ccc} \lambda_1  & 0 & 0 \\ 0 & \lambda_2 & 0 \\ 0 & 0 & \lambda_3 \end{array} \right). After you apply rotation by U to the curves (a)” ~ (i)”, those curves are symmetrically placed about the xyz axes, and their center points still cross the origin. The resulting curves look like below. Or rather I should say you projected (a)’ ~ (i)’ on their eigenvectors.

In this article mainly (a)” , (g)”, (h)”, and (i)” are important. General equations for the curves is as follows

  • (a)”: \frac{x^2}{l^2} + \frac{y^2}{m^2} + \frac{z^2}{n^2} = 1
  • (g)”: z = \frac{x^2}{l^2} + \frac{y^2}{m^2}
  • (h)”: z = \frac{x^2}{l^2} - \frac{y^2}{m^2}
  • (i)”: z = \frac{x^2}{l^2}

, where l, m, n \in \mathbb{R}^+.

Even if this section has been puzzling to you, you just have to keep one point in your mind: we have been discussing general quadratic curves, but in PCA, you only need to consider a case where A is a covariance matrix, that is A=\Sigma. PCA corresponds to the case where you shift and rotate the curve (a) into (a)”. Subtracting the mean of data from each point of data corresponds to shifting quadratic curve (a) to (a)’. Calculating eigenvectors of A corresponds to calculating a rotation matrix U such that the curve (a)’ comes to (a)” after applying the rotation, or projecting curves on eigenvectors of \Sigma. Importantly we are only discussing the covariance of certain data, not the distribution of the data itself.

*Just in case you are interested in a little more mathematical sides: it is known that if you rotate all the points \boldsymbol{x} on the curve \boldsymbol{x}^T A\boldsymbol{x} + 2\boldsymbol{b}^T\boldsymbol{x} + s = 0 with the rotation matrix P, those points \boldsymbol{x} are mapped into a new quadratic curve \alpha x^2 + \beta y^2 + \gamma z^2 + \lambda x + \mu y + \nu z + \rho = 0. That means the rotation of the original quadratic curve with P (or rather rotating axes) enables getting rid of the terms xy, yz, zx. Also it is known that when \alpha ' \neq 0, with proper translations and rotations, the quadratic curve \alpha x^2 + \beta y^2 + \gamma z^2 + \lambda x + \mu y + \nu z + \rho = 0 can be mapped into one of the types of quadratic curves in the figure below, depending on coefficients of the original quadratic curve. And the discussion so far can be generalized to higher dimensional spaces, but that is beyond the scope of this article series. Please consult decent textbooks on linear algebra around you for further details.

4 Eigenvectors are gradients and sometimes variances.

In the second section I explained that you can express quadratic functions f(\boldsymbol{x}; A) = \boldsymbol{x}^T A \boldsymbol{x} in a very simple way by projecting \boldsymbol{x} on eigenvectors of A.

You can comprehend what I have explained in another way: eigenvectors, to be exact eigenvectors of real symmetric matrices A, are gradients. And in case of PCA, I mean when A=\Sigma eigenvalues are also variances. Before explaining what that means, let me explain a little of the totally common facts on mathematics. If you have variables \boldsymbol{x}\in \mathbb{R}^D, I think you can comprehend functions f(\boldysmbol{x}) in two ways. One is a normal “functions” f(\boldsymbol{x}), and the others are “curves” f(\boldsymbol{x}) = const.. “Functions” get an input \boldsymbol{x} and gives out an output f(\boldsymbol{x}), just as well as normal functions you would imagine. “Curves” are rather sets of \boldsymbol{x} \in \mathbb{R}^D such that f(\boldsymbol{x}) = const..

*Please assume that the terms “functions” and “curves” are my original words. I use them just in case I fail to use functions and curves properly.

The quadratic curves in the figure above are all “curves” in my term, which can be denoted as f(\boldsymbol{x}; A_3, \boldsymbol{b}_3)=const or f(\boldsymbol{x}; A_3)=const. However if you replace z of (g)”, (h)”, and (i)” with f, you can interpret the “curves” as “functions” which are denoted as f(\boldsymbol{x}; A_2). This might sounds too obvious to you, and my point is you can visualize how values of “functions” change only when the inputs are 2 dimensional.

When a symmetric 2\times 2 real matrices A_2 have two eigenvalues \lambda_1, \lambda_2, the distribution of quadratic curves can be roughly classified to the following three types.

  • (g): Both \lambda_1 and \lambda_2 are positive or negative.
  • (h): Either of \lambda_1 or \lambda_2 is positive and the other is negative.
  • (i): Either of \lambda_1 or \lambda_2 is 0 and the other is not.

The equations of (g)” , (h)”, and (i)” correspond to each type of f=(\boldsymbol{x}; A_2), and thier curves look like the three graphs below.

And in fact, when start from the origin and go in the direction of an eigenvector \boldsymbol{u}_i, \lambda_i is the gradient of the direction. You can see that more clearly when you restrict the distribution of f=(\boldsymbol{x}; A_2) to a unit circle. Like in the figure below, in case \lambda_1 = 7, \lambda_2 = 3, which is classified to (g), the distribution looks like the left side, and if you restrict the distribution in the unit circle, the distribution looks like a bowl like the middle and the right side. When you move in the direction of \boldsymbol{u}_1, you can climb the bowl as as high as \lambda_1, in \boldsymbol{u}_2 as high as \lambda_2.

Also in case of (h), the same facts hold. But in this case, you can also descend the curve.

*You might have seen the curve above in the context of optimization with stochastic gradient descent. The origin of the curve above is a notorious saddle point, where gradients are all 0 in any directions but not a local maximum or minimum. Points can be stuck in this point during optimization.

Especially in case of PCA, A is a covariance matrix, thus A=\Sigma. Eigenvalues of \Sigma are all equal to or greater than 0. And it is known that in this case \lambda_i is the variance of data projected on its corresponding eigenvector \boldsymbol{u}_i (i=0, \dots , D). Hence, if you project f(\boldsymbol{x}; \Sigma), quadratic curves formed by a covariance matrix \Sigma, on eigenvectors of \Sigma, you get f(\boldsymbol{x}; \Sigma) = ({x'}_1 \: \dots \: {x'}_D) (\lambda_1 {x'}_1 \: \dots \: \lambda_D {x'}_D)^t =\lambda_1 ({x'}_1)^2 + \cdots + \lambda_D ({x'}_D)^2.  This shows that you can re-weight ({x'}_1 \: \dots \: {x'}_D), the coordinates of data projected projected on eigenvectors of A, with \lambda_1, \dots, \lambda_D, which are variances ({x'}_1 \: \dots \: {x'}_D). As I mentioned in an example of data of exam scores in the last article, the bigger a variance \lambda_i is, the more the feature described by \boldsymbol{u}_i vary from sample to sample. In other words, you can ignore eigenvectors corresponding to small eigenvalues.

That is a great hint why principal components corresponding to large eigenvectors contain much information of the data distribution. And you can also interpret PCA as a “climbing” a bowl of f(\boldsymbol{x}; A_D), as I have visualized in the case of (g) type curve in the figure above.

*But as I have repeatedly mentioned, ellipsoid which fit data well isf(\boldsymbol{x}; \Sigma ^{-1}) =(\boldsymbol{x}')^T diag(\frac{1}{\lambda_1}, \dots, \frac{1}{\lambda_D})\boldsymbol{x}' = \frac{({x'}_{1})^2}{\lambda_1} + \cdots + \frac{({x'}_{D})^2}{\lambda_D} = const..

*You have to be careful that even if you slice a type (h) curve f(\boldsymbol{x}; A_D) with a place z=const. the resulting cross section does not fit the original data well because the equation of the cross section is \lambda_1 ({x'}_1)^2 + \cdots + \lambda_D ({x'}_D)^2 = const. The figure below is an example of slicing the same f(\boldsymbol{x}; A_2) as the one above with z=1, and the resulting cross section.

As we have seen, \lambda_i, the eigenvalues of the covariance matrix of data are variances or data when projected on it eigenvectors. At the same time, when you fit an ellipsoid on the data, \sqrt{\lambda_i} is the radius of the ellipsoid corresponding to \boldsymbol{u}_i. Thus ignoring data projected on eigenvectors corresponding to small eigenvalues is equivalent to cutting of the axes of the ellipsoid with small radiusses.

I have explained PCA in three different ways over three articles.

  • The second article: I focused on what kind of linear transformations convariance matrices \Sigma enable, by visualizing displacement vectors. And those vectors look like swirling and extending into directions of eigenvectors of \Sigma.
  • The third article: We directly found directions where certain data distribution “swell” the most, to find that data swell the most in directions of eigenvectors.
  • In this article, we have seen PCA corresponds to only one case of quadratic functions, where the matrix A is a covariance matrix. When you go in the directions of eigenvectors corresponding to big eigenvalues, the quadratic function increases the most. Also that means data samples have bigger variances when projected on the eigenvectors. Thus you can cut off eigenvectors corresponding to small eigenvectors because they retain little information about data, and that is equivalent to fitting an ellipsoid on data and cutting off axes with small radiuses.

*Let A be a covariance matrix, and you can diagonalize it with an orthogonal matrix U as follow: U^{T}AU = \Lambda, where \Lambda = diag(\lambda_1, \dots, \lambda_D). Thus A = U \Lambda U^{T}. U is a rotation, and multiplying a \boldsymbol{x} with \Lambda means you multiply each eigenvalue to each element of \boldsymbol{x}. At the end U^T enables the reverse rotation.

If you get data like the left side of the figure below, most explanation on PCA would just fit an oval on this data distribution. However after reading this articles series so far, you would have learned to see PCA from different viewpoints like at the right side of the figure below.

 

5 Ellipsoids in Gaussian distributions.

I have explained that if the covariance of a data distribution is \boldsymbol{\Sigma}, the ellipsoid which fits the distribution the best is \bigl((\boldsymbol{x} - \boldsymbol{\mu}), \boldsymbol{\Sigma}^{-1}(\boldsymbol{x} - \boldsymbol{\mu})\bigr) = 1. You might have seen the part \bigl((\boldsymbol{x} - \boldsymbol{\mu}), \boldsymbol{\Sigma}^{-1}(\boldsymbol{x} - \boldsymbol{\mu})\bigr) = (\boldsymbol{x} - \boldsymbol{\mu}) \boldsymbol{\Sigma}^{-1}(\boldsymbol{x} - \boldsymbol{\mu}) somewhere else. It is the exponent of general Gaussian distributions: \mathcal{N}(\boldsymbol{x} | \boldsymbol{\mu}, \boldsymbol{\Sigma}) = \frac{1}{(2\pi)^{D/2}} \frac{1}{|\boldsymbol{\Sigma}|} exp\{ -\frac{1}{2}(\boldsymbol{x} - \boldsymbol{\mu}) \boldsymbol{\Sigma}^{-1}(\boldsymbol{x} - \boldsymbol{\mu}) \}.  It is known that the eigenvalues of \Sigma ^{-1} are \frac{1}{\lambda_1}, \dots, \frac{1}{\lambda_D}, and eigenvectors corresponding to each eigenvalue are also \boldsymbol{u}_1, \dots, \boldsymbol{u}_D respectively. Hence just as well as what we have seen, if you project (\boldsymbol{x} - \boldsymbol{\mu}) on each eigenvector of \Sigma ^{-1}, we can convert the exponent of the Gaussian distribution.

Let -\frac{1}{2}(\boldsymbol{x} - \boldsymbol{\mu}) \boldsymbol{\Sigma}^{-1}(\boldsymbol{x} - \boldsymbol{\mu}) be \boldsymbol{y} and U ^{-1} \boldsymbol{y}= U^{T} \boldsymbol{y} be \boldsymbol{y}', where U=(\boldsymbol{u}_1 \: \dots \: \boldsymbol{u}_D). Just as we have seen, (\boldsymbol{x} - \boldsymbol{\mu}) \boldsymbol{\Sigma}^{-1}(\boldsymbol{x} - \boldsymbol{\mu}) =\boldsymbol{y}^T\Sigma^{-1} \boldsymbol{y} =(U\boldsymbol{y}')^T \Sigma^{-1} U\boldsymbol{y}' =((\boldsymbol{y}')^T U^T \Sigma^{-1} U\boldsymbol{y}' = (\boldsymbol{y}')^T diag(\frac{1}{\lambda_1}, \dots, \frac{1}{\lambda_D}) \boldsymbol{y}' = \frac{({y'}_{1})^2}{\lambda_1} + \cdots + \frac{({y'}_{D})^2}{\lambda_D}. Hence \mathcal{N}(\boldsymbol{x} | \boldsymbol{\mu}, \boldsymbol{\Sigma}) = \frac{1}{(2\pi)^{D/2}} \frac{1}{|\boldsymbol{\Sigma}|} exp\{ -\frac{1}{2}(\boldsymbol{y}) \boldsymbol{\Sigma}^{-1}(\boldsymbol{y}) \} =  \frac{1}{(2\pi)^{D/2}} \frac{1}{|\boldsymbol{\Sigma}|} exp\{ -\frac{1}{2}(\frac{({y'}_{1})^2}{\lambda_1} + \cdots + \frac{({y'}_{D})^2}{\lambda_D} ) \} =\frac{1}{(2\pi)^{1/2}} \frac{1}{|\boldsymbol{\Sigma}|} exp\biggl( -\frac{1}{2} \frac{({y'}_{1})^2}{\lambda_1} \biggl) \cdots \frac{1}{(2\pi)^{1/2}} \frac{1}{|\boldsymbol{\Sigma}|} exp\biggl( -\frac{1}{2}\frac{({y'}_{D})^2}{\lambda_D} \biggl).

*To be mathematically exact about changing variants of normal distributions, you have to consider for example Jacobian matrices.

This results above demonstrate that, by projecting data on the eigenvectors of its covariance matrix, you can factorize the original multi-dimensional Gaussian distribution into a product of Gaussian distributions which are irrelevant to each other. However, at the same time, that is the potential limit of approximating data with PCA. This idea is going to be more important when you think about more probabilistic ways to handle PCA, which is more robust to lack of data.

I have explained PCA over 3 articles from various viewpoints. If you have been patient enough to read my article series, I think you have gained some deeper insight into not only PCA, but also linear algebra, and that should be helpful when you learn or teach data science. I hope my codes also help you. In fact these are not the only topics about PCA. There are a lot of important PCA-like algorithms.

In fact our expedition of ellipsoids, or PCA still continues, just as Star Wars series still continues. Especially if I have to explain an algorithm named probabilistic PCA, I need to explain the “Bayesian world” of machine learning. Most machine learning algorithms covered by major introductory textbooks tend to be too deterministic and dependent on the size of data. Many of those algorithms have another “parallel world,” where you can handle inaccuracy in better ways. I hope I can also write about them, and I might prepare another trilogy for such PCA. But I will not disappoint you, like “The Phantom Menace.”

Appendix: making a model of a bunch of grape with ellipsoid berries.

If you can control quadratic curves, reshaping and rotating them, you can make a model of a grape of olive bunch on Matplotlib. I made a program of making a model of a bunch of berries on Matplotlib using the module to draw ellipsoids which I introduced earlier. You can check the codes in this page.

*I have no idea how many people on this earth are in need of making such models.

I made some modules so that you can see the grape bunch from several angles. This might look very simple to you, but the locations of berries are organized carefully so that it looks like they are placed around a stem and that the berries are not too close to each other.

 

The programming code I created for this article is completly available here.

[Refereces]

[1]C. M. Bishop, “Pattern Recognition and Machine Learning,” (2006), Springer, pp. 78-83, 559-577

[2]「理工系新課程 線形代数 基礎から応用まで」, 培風館、(2017)

[3]「これなら分かる 最適化数学 基礎原理から計算手法まで」, 金谷健一著、共立出版, (2019), pp. 17-49

[4]「これなら分かる 応用数学教室 最小二乗法からウェーブレットまで」, 金谷健一著、共立出版, (2019), pp.165-208

[5] 「サボテンパイソン 」
https://sabopy.com/

 

Multi-head attention mechanism: “queries”, “keys”, and “values,” over and over again

This is the third article of my article series named “Instructions on Transformer for people outside NLP field, but with examples of NLP.”

In the last article, I explained how attention mechanism works in simple seq2seq models with RNNs, and it basically calculates correspondences of the hidden state at every time step, with all the outputs of the encoder. However I would say the attention mechanisms of RNN seq2seq models use only one standard for comparing them. Using only one standard is not enough for understanding languages, especially when you learn a foreign language. You would sometimes find it difficult to explain how to translate a word in your language to another language. Even if a pair of languages are very similar to each other, translating them cannot be simple switching of vocabulary. Usually a single token in one language is related to several tokens in the other language, and vice versa. How they correspond to each other depends on several criteria, for example “what”, “who”, “when”, “where”, “why”, and “how”. It is easy to imagine that you should compare tokens with several criteria.

Transformer model was first introduced in the original paper named “Attention Is All You Need,” and from the title you can easily see that attention mechanism plays important roles in this model. When you learn about Transformer model, you will see the figure below, which is used in the original paper on Transformer.  This is the simplified overall structure of one layer of Transformer model, and you stack this layer N times. In one layer of Transformer, there are three multi-head attention, which are displayed as boxes in orange. These are the very parts which compare the tokens on several standards. I made the head article of this article series inspired by this multi-head attention mechanism.

The figure below is also from the original paper on Transfromer. If you can understand how multi-head attention mechanism works with the explanations in the paper, and if you have no troubles understanding the codes in the official Tensorflow tutorial, I have to say this article is not for you. However I bet that is not true of majority of people, and at least I need one article to clearly explain how multi-head attention works. Please keep it in mind that this article covers only the architectures of the two figures below. However multi-head attention mechanisms are crucial components of Transformer model, and throughout this article, you would not only see how they work but also get a little control over it at an implementation level.

1 Multi-head attention mechanism

When you learn Transformer model, I recommend you first to pay attention to multi-head attention. And when you learn multi-head attentions, before seeing what scaled dot-product attention is, you should understand the whole structure of multi-head attention, which is at the right side of the figure above. In order to calculate attentions with a “query”, as I said in the last article, “you compare the ‘query’ with the ‘keys’ and get scores/weights for the ‘values.’ Each score/weight is in short the relevance between the ‘query’ and each ‘key’. And you reweight the ‘values’ with the scores/weights, and take the summation of the reweighted ‘values’.” Sooner or later, you will notice I would be just repeating these phrases over and over again throughout this article, in several ways.

*Even if you are not sure what “reweighting” means in this context, please keep reading. I think you would little by little see what it means especially in the next section.

The overall process of calculating multi-head attention, displayed in the figure above, is as follows (Please just keep reading. Please do not think too much.): first you split the V: “values”, K: “keys”, and Q: “queries”, and second you transform those divided “values”, “keys”, and “queries” with densely connected layers (“Linear” in the figure). Next you calculate attention weights and reweight the “values” and take the summation of the reiweighted “values”, and you concatenate the resulting summations. At the end you pass the concatenated “values” through another densely connected layers. The mechanism of scaled dot-product attention is just a matter of how to concretely calculate those attentions and reweight the “values”.

*In the last article I briefly mentioned that “keys” and “queries” can be in the same language. They can even be the same sentence in the same language, and in this case the resulting attentions are called self-attentions, which we are mainly going to see. I think most people calculate “self-attentions” unconsciously when they speak. You constantly care about what “she”, “it” , “the”, or “that” refers to in you own sentence, and we can say self-attention is how these everyday processes is implemented.

Let’s see the whole process of calculating multi-head attention at a little abstract level. From now on, we consider an example of calculating multi-head self-attentions, where the input is a sentence “Anthony Hopkins admired Michael Bay as a great director.” In this example, the number of tokens is 9, and each token is encoded as a 512-dimensional embedding vector. And the number of heads is 8. In this case, as you can see in the figure below, the input sentence “Anthony Hopkins admired Michael Bay as a great director.” is implemented as a 9\times 512 matrix. You first split each token into 512/8=64 dimensional, 8 vectors in total, as I colored in the figure below. In other words, the input matrix is divided into 8 colored chunks, which are all 9\times 64 matrices, but each colored matrix expresses the same sentence. And you calculate self-attentions of the input sentence independently in the 8 heads, and you reweight the “values” according to the attentions/weights. After this, you stack the sum of the reweighted “values”  in each colored head, and you concatenate the stacked tokens of each colored head. The size of each colored chunk does not change even after reweighting the tokens. According to Ashish Vaswani, who invented Transformer model, each head compare “queries” and “keys” on each standard. If the a Transformer model has 4 layers with 8-head multi-head attention , at least its encoder has 4\times 8 = 32 heads, so the encoder learn the relations of tokens of the input on 32 different standards.

I think you now have rough insight into how you calculate multi-head attentions. In the next section I am going to explain the process of reweighting the tokens, that is, I am finally going to explain what those colorful lines in the head image of this article series are.

*Each head is randomly initialized, so they learn to compare tokens with different criteria. The standards might be straightforward like “what” or “who”, or maybe much more complicated. In attention mechanisms in deep learning, you do not need feature engineering for setting such standards.

2 Calculating attentions and reweighting “values”

If you have read the last article or if you understand attention mechanism to some extent, you should already know that attention mechanism calculates attentions, or relevance between “queries” and “keys.” In the last article, I showed the idea of weights as a histogram, and in that case the “query” was the hidden state of the decoder at every time step, whereas the “keys” were the outputs of the encoder. In this section, I am going to explain attention mechanism in a more abstract way, and we consider comparing more general “tokens”, rather than concrete outputs of certain networks. In this section each [ \cdots ] denotes a token, which is usually an embedding vector in practice.

Please remember this mantra of attention mechanism: “you compare the ‘query’ with the ‘keys’ and get scores/weights for the ‘values.’ Each score/weight is in short the relevance between the ‘query’ and each ‘key’. And you reweight the ‘values’ with the scores/weights, and take the summation of the reweighted ‘values’.” The figure below shows an overview of a case where “Michael” is a query. In this case you compare the query with the “keys”, that is, the input sentence “Anthony Hopkins admired Michael Bay as a great director.” and you get the histogram of attentions/weights. Importantly the sum of the weights 1. With the attentions you have just calculated, you can reweight the “values,” which also denote the same input sentence. After that you can finally take a summation of the reweighted values. And you use this summation.

*I have been repeating the phrase “reweighting ‘values’  with attentions,”  but you in practice calculate the sum of those reweighted “values.”

Assume that compared to the “query”  token “Michael”, the weights of the “key” tokens “Anthony”, “Hopkins”, “admired”, “Michael”, “Bay”, “as”, “a”, “great”, and “director.” are respectively 0.06, 0.09, 0.05, 0.25, 0.18, 0.06, 0.09, 0.06, 0.15. In this case the sum of the reweighted token is 0.06″Anthony” + 0.09″Hopkins” + 0.05″admired” + 0.25″Michael” + 0.18″Bay” + 0.06″as” + 0.09″a” + 0.06″great” 0.15″director.”, and this sum is the what wee actually use.

*Of course the tokens are embedding vectors in practice. You calculate the reweighted vector in actual implementation.

You repeat this process for all the “queries.”  As you can see in the figure below, you get summations of 9 pairs of reweighted “values” because you use every token of the input sentence “Anthony Hopkins admired Michael Bay as a great director.” as a “query.” You stack the sum of reweighted “values” like the matrix in purple in the figure below, and this is the output of a one head multi-head attention.

3 Scaled-dot product

This section is a only a matter of linear algebra. Maybe this is not even so sophisticated as linear algebra. You just have to do lots of Excel-like operations. A tutorial on Transformer by Jay Alammar is also a very nice study material to understand this topic with simpler examples. I tried my best so that you can clearly understand multi-head attention at a more mathematical level, and all you need to know in order to read this section is how to calculate products of matrices or vectors, which you would see in the first some pages of textbooks on linear algebra.

We have seen that in order to calculate multi-head attentions, we prepare 8 pairs of “queries”, “keys” , and “values”, which I showed in 8 different colors in the figure in the first section. We calculate attentions and reweight “values” independently in 8 different heads, and in each head the reweighted “values” are calculated with this very simple formula of scaled dot-product: Attention(\boldsymbol{Q}, \boldsymbol{K}, \boldsymbol{V}) =softmax(\frac{\boldsymbol{Q} \boldsymbol{K} ^T}{\sqrt{d}_k})\boldsymbol{V}. Let’s take an example of calculating a scaled dot-product in the blue head.

At the left side of the figure below is a figure from the original paper on Transformer, which explains one-head of multi-head attention. If you have read through this article so far, the figure at the right side would be more straightforward to understand. You divide the input sentence into 8 chunks of matrices, and you independently put those chunks into eight head. In one head, you convert the input matrix by three different fully connected layers, which is “Linear” in the figure below, and prepare three matrices Q, K, V, which are “queries”, “keys”, and “values” respectively.

*Whichever color attention heads are in, the processes are all the same.

*You divide \frac{\boldsymbol{Q}} {\boldsymbol{K}^T} by \sqrt{d}_k in the formula. According to the original paper, it is known that re-scaling \frac{\boldsymbol{Q} }{\boldsymbol{K}^T} by \sqrt{d}_k is found to be effective. I am not going to discuss why in this article.

As you can see in the figure below, calculating Attention(\boldsymbol{Q}, \boldsymbol{K}, \boldsymbol{V}) is virtually just multiplying three matrices with the same size (Only K is transposed though). The resulting 9\times 64 matrix is the output of the head.

softmax(\frac{\boldsymbol{Q} \boldsymbol{K} ^T}{\sqrt{d}_k}) is calculated like in the figure below. The softmax function regularize each row of the re-scaled product \frac{\boldsymbol{Q} \boldsymbol{K} ^T}{\sqrt{d}_k}, and the resulting 9\times 9 matrix is a kind a heat map of self-attentions.

The process of comparing one “query” with “keys” is done with simple multiplication of a vector and a matrix, as you can see in the figure below. You can get a histogram of attentions for each query, and the resulting 9 dimensional vector is a list of attentions/weights, which is a list of blue circles in the figure below. That means, in Transformer model, you can compare a “query” and a “key” only by calculating an inner product. After re-scaling the vectors by dividing them with \sqrt{d_k} and regularizing them with a softmax function, you stack those vectors, and the stacked vectors is the heat map of attentions.

You can reweight “values” with the heat map of self-attentions, with simple multiplication. It would be more straightforward if you consider a transposed scaled dot-product \boldsymbol{V}^T \cdot softmax(\frac{\boldsymbol{Q} \boldsymbol{K} ^T}{\sqrt{d}_k})^T. This also should be easy to understand if you know basics of linear algebra.

One column of the resulting matrix (\boldsymbol{V}^T \cdot softmax(\frac{\boldsymbol{Q} \boldsymbol{K} ^T}{\sqrt{d}_k})^T) can be calculated with a simple multiplication of a matrix and a vector, as you can see in the figure below. This corresponds to the process or “taking a summation of reweighted ‘values’,” which I have been repeating. And I would like you to remember that you got those weights (blue) circles by comparing a “query” with “keys.”

Again and again, let’s repeat the mantra of attention mechanism together: “you compare the ‘query’ with the ‘keys’ and get scores/weights for the ‘values.’ Each score/weight is in short the relevance between the ‘query’ and each ‘key’. And you reweight the ‘values’ with the scores/weights, and take the summation of the reweighted ‘values’.” If you have been patient enough to follow my explanations, I bet you have got a clear view on how multi-head attention mechanism works.

We have been seeing the case of the blue head, but you can do exactly the same procedures in every head, at the same time, and this is what enables parallelization of multi-head attention mechanism. You concatenate the outputs of all the heads, and you put the concatenated matrix through a fully connected layers.

If you are reading this article from the beginning, I think this section is also showing the same idea which I have repeated, and I bet more or less you no have clearer views on how multi-head attention mechanism works. In the next section we are going to see how this is implemented.

4 Tensorflow implementation of multi-head attention

Let’s see how multi-head attention is implemented in the Tensorflow official tutorial. If you have read through this article so far, this should not be so difficult. I also added codes for displaying heat maps of self attentions. With the codes in this Github page, you can display self-attention heat maps for any input sentences in English.

The multi-head attention mechanism is implemented as below. If you understand Python codes and Tensorflow to some extent, I think this part is relatively easy.  The multi-head attention part is implemented as a class because you need to train weights of some fully connected layers. Whereas, scaled dot-product is just a function.

*I am going to explain the create_padding_mask() and create_look_ahead_mask() functions in upcoming articles. You do not need them this time.

Let’s see a case of using multi-head attention mechanism on a (1, 9, 512) sized input tensor, just as we have been considering in throughout this article. The first axis of (1, 9, 512) corresponds to the batch size, so this tensor is virtually a (9, 512) sized tensor, and this means the input is composed of 9 512-dimensional vectors. In the results below, you can see how the shape of input tensor changes after each procedure of calculating multi-head attention. Also you can see that the output of the multi-head attention is the same as the input, and you get a 9\times 9 matrix of attention heat maps of each attention head.

I guess the most complicated part of this implementation above is the split_head() function, especially if you do not understand tensor arithmetic. This part corresponds to splitting the input tensor to 8 different colored matrices as in one of the figures above. If you cannot understand what is going on in the function, I recommend you to prepare a sample tensor as below.

This is just a simple (1, 9, 512) sized tensor with sequential integer elements. The first row (1, 2, …., 512) corresponds to the first input token, and (4097, 4098, … , 4608) to the last one. You should try converting this sample tensor to see how multi-head attention is implemented. For example you can try the operations below.

These operations correspond to splitting the input into 8 heads, whose sizes are all (9, 64). And the second axis of the resulting (1, 8, 9, 64) tensor corresponds to the index of the heads. Thus sample_sentence[0][0] corresponds to the first head, the blue 9\times 64 matrix. Some Tensorflow functions enable linear calculations in each attention head, independently as in the codes below.

Very importantly, we have been only considering the cases of calculating self attentions, where all “queries”, “keys”, and “values” come from the same sentence in the same language. However, as I showed in the last article, usually “queries” are in a different language from “keys” and “values” in translation tasks, and “keys” and “values” are in the same language. And as you can imagine, usualy “queries” have different number of tokens from “keys” or “values.” You also need to understand this case, which is not calculating self-attentions. If you have followed this article so far, this case is not that hard to you. Let’s briefly see an example where the input sentence in the source language is composed 9 tokens, on the other hand the output is composed 12 tokens.

As I mentioned, one of the outputs of each multi-head attention class is 9\times 9 matrix of attention heat maps, which I displayed as a matrix composed of blue circles in the last section. The the implementation in the Tensorflow official tutorial, I have added codes to display actual heat maps of any input sentences in English.

*If you want to try displaying them by yourself, download or just copy and paste codes in this Github page. Please maker “datasets” directory in the same directory as the code. Please download “spa-eng.zip” from this page, and unzip it. After that please put “spa.txt” on the “datasets” directory. Also, please download the “checkpoints_en_es” folder from this link, and place the folder in the same directory as the file in the Github page. In the upcoming articles, you would need similar processes to run my codes.

After running codes in the Github page, you can display heat maps of self attentions. Let’s input the sentence “Anthony Hopkins admired Michael Bay as a great director.” You would get a heat maps like this.

In fact, my toy implementation cannot handle proper nouns such as “Anthony” or “Michael.” Then let’s consider a simple input sentence “He admired her as a great director.” In each layer, you respectively get 8 self-attention heat maps.

I think we can see some tendencies in those heat maps. The heat maps in the early layers, which are close to the input, are blurry. And the distributions of the heat maps come to concentrate more or less diagonally. At the end, presumably they learn to pay attention to the start and the end of sentences.

You have finally finished reading this article. Congratulations.

You should be proud of having been patient, and you passed the most tiresome part of learning Transformer model. You must be ready for making a toy English-German translator in the upcoming articles. Also I am sure you have understood that Michael Bay is a great director, no matter what people say.

*Hannibal Lecter, I mean Athony Hopkins, also wrote a letter to the staff of “Breaking Bad,” and he told them the tv show let him regain his passion. He is a kind of admiring around, and I am a little worried that he might be getting senile. He played a role of a father forgetting his daughter in his new film “The Father.” I must see it to check if that is really an acting, or not.

[References]

[1] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, Illia Polosukhin, “Attention Is All You Need” (2017)

[2] “Transformer model for language understanding,” Tensorflow Core
https://www.tensorflow.org/overview

[3] “Neural machine translation with attention,” Tensorflow Core
https://www.tensorflow.org/tutorials/text/nmt_with_attention

[4] Jay Alammar, “The Illustrated Transformer,”
http://jalammar.github.io/illustrated-transformer/

[5] “Stanford CS224N: NLP with Deep Learning | Winter 2019 | Lecture 14 – Transformers and Self-Attention,” stanfordonline, (2019)
https://www.youtube.com/watch?v=5vcj8kSwBCY

[6]Tsuboi Yuuta, Unno Yuuya, Suzuki Jun, “Machine Learning Professional Series: Natural Language Processing with Deep Learning,” (2017), pp. 91-94
坪井祐太、海野裕也、鈴木潤 著, 「機械学習プロフェッショナルシリーズ 深層学習による自然言語処理」, (2017), pp. 191-193

[7]”Stanford CS224N: NLP with Deep Learning | Winter 2019 | Lecture 8 – Translation, Seq2Seq, Attention”, stanfordonline, (2019)
https://www.youtube.com/watch?v=XXtpJxZBa2c

[8]Rosemary Rossi, “Anthony Hopkins Compares ‘Genius’ Michael Bay to Spielberg, Scorsese,” yahoo! entertainment, (2017)
https://www.yahoo.com/entertainment/anthony-hopkins-transformers-director-michael-bay-guy-genius-010058439.html

* I make study materials on machine learning, sponsored by DATANOMIQ. I do my best to make my content as straightforward but as precise as possible. I include all of my reference sources. If you notice any mistakes in my materials, including grammatical errors, please let me know (email: yasuto.tamura@datanomiq.de). And if you have any advice for making my materials more understandable to learners, I would appreciate hearing it.

Data Mining Process flow – Easy Understanding

1 Overview

Development of computer processing power, network and automated software completely change and give new concept of each business. And data mining play the vital part to solve, finding the hidden patterns and relationship from large dataset with business by using sophisticated data analysis tools like methodology, method, process flow etc.

On this paper, proposed a process flow followed CRISP-DM methodology and has six steps where data understanding does not considered.

Phase of new process flow given below:-

Phase 1: Involved with collection, outliner treatment, imputation, transformation, scaling, and partition dataset in to two sub-frames (Training and Testing). Here as an example for outliner treatment, imputation, transformation, scaling consider accordingly Z score, mean, One hot encoding and Min Max Scaler.

Phase 2: On this Phase training and testing data balance with same balancing algorithm but separately. As an example here SMOTE (synthetic minority oversampling technique) is considered.

Phase 3: This phase involved with reduction, selection, aggregation, extraction. But here for an example considering same feature reduction algorithm (LDA -Linear Discriminant analysis) on training and testing data set separately.

Phase 4: On this Phase Training data set again partition into two more set (Training and Validation).

Phase 5: This Phase considering several base algorithms as a base model like CNN, RNN, Random forest, MLP, Regression, Ensemble method. This phase also involve to find out best hyper parameter and sub-algorithm for each base algorithm. As an example on this paper consider two class classification problems and also consider Random forest (Included CART – Classification and Regression Tree and GINI index impurity) and MLP classifier (Included (Relu, Sigmoid, binary cross entropy, Adam – Adaptive Moment Estimation) as base algorithms.

Phase 6: First, Prediction with validation data then evaluates with Test dataset which is fully unknown for these (Random forest, MLP classifier) two base algorithms. Then calculate the confusion matrix, ROC, AUC to find the best base algorithm.

New method from phase 1 to phase 4 followed CRISP-DM methodology steps such as data collection, data preparation then phase 5 followed modelling and phase 6 followed evaluation and implementation steps.

Structure of proposed process flow for two class problem combined with algorithm and sub-algorithm display on figure – 1.

These articles mainly focus to describe all algorithms which are going to implementation for better understanding.

 

 

Data Mining Process Flow

Figure 1 – Data Mining Process Flow

2 Phase 1: Outlier treatment, Transform, Scaling, Imputation

This phase involved with outlier treatment, imputation, scaling, and transform data.

2.1 Outliner treatment: – Z score

Outlier is a data point which lies far from all other data point in a data set. Outlier need to treat because it may bias the entire result. Outlier treatment with Z score is a common technique.  Z score is a standard score in statistics.  Z score provides information about data value is smaller or grater then mean that means how many standard deviations away from the mean value. Z score equation display below:

Z = \frac{(x - \mu)}{\sigma}

Here x = data point
σ = Standard deviation
μ = mean value

Equation- 1 Z-Score

In a normal distribution Z score represent 68% data lies on +/- 1, 95% data point lies on +/- 2, 99.7% data point lies on +/- 3 standard deviation.

2.2 Imputation data: – mean

Imputation is a way to handle missing data by replacing substituted value. There are many imputation technique represent like mean, median, mode, k-nearest neighbours. Mean imputation is the technique to replacing missing information with mean value. On the mean imputation first calculate the particular features mean value and then replace the missing value with mean value. The next equation displays the mean calculation:

\mu = \frac{(\sum x)}{n}

Here x = value of each point
n = number of values
μ = mean value

Equation- 2 Mean

2.3 Transform: – One hot encoding

Encoding is a pre-processing technique which represents data in such a way that computer can understand.  For understanding of machine learning algorithm categorical columns convert to numerical columns, this process called categorical encoding. There are multiple way to handle categorical variable but most widely used techniques are label encoding and one host encoding. On label encoding give a numeric (integer number) for each category. Suppose there are 3 categories of foods like apples, orange, banana. When label encoding is used then 3 categories will get a numerical value like apples = 1, banana = 2 and orange = 3. But there is very high probability that machine learning model can capture the relationship in between categories such as apple < banana < orange or calculate average across categories like 1 +3 = 4 / 2 = 2 that means model can understand average of apple and orange together is banana which is not acceptable because model correlation calculation is wrong. For solving this problem one hot encoding appear. The following table displays the label encoding is transformed into one hot encoding.

Label Encoding and One-Hot-Encoding

Table- 1 Encoding example

On hot encoding categorical value split into columns and each column contains 0 or 1 according to columns placement.

2.4 Scaling data: – Min Max Scaler

Feature scaling method is standardized or normalization the independent variable that means it is used to scale the data in a particular range like -1 to +1 or depending on algorithm. Generally normalization used where data distribution does not follow Gaussian distribution and standardization used where data distribution follow Gaussian distribution. On standardization techniques transform data values are cantered around the mean and unit is standard deviation. Formula for standardization given below:

Standardization X = \frac{(X - \mu)}{\sigma}

Equation-3 Equations for Standardization

X represent the feature value, µ represent mean of the feature value and σ represent standard deviation of the feature value. Standardized data value does not restrict to a particular range.

Normalization techniques shifted and rescaled data value range between 0 and 1. Normalization techniques also called Min-Max scaling. Formula for normalization given below:

Normalization X = \frac{(X - X_{min})}{X_{max} - X_{min}}

Equation – 4 Equations for Normalization

Above X, Xmin, Xmax are accordingly feature values, feature minimum value and feature maximum value. On above formula when X is minimums then numerator will be 0 (  is 0) or if X is maximums then the numerator is equal to the denominator (  is 1). But when X data value between minimum and maximum then  is between 0 and 1. If ranges value of data does not normalized then bigger range can influence the result.

3 Phase 2: – Balance Data

3.1 SMOTE

SMOTE (synthetic minority oversampling technique) is an oversampling technique where synthetic observations are created based on existing minority observations. This technique operates in feature space instead of data space. Under SMOTE each minority class observation calculates k nearest neighbours and randomly chose the neighbours depending on over-sampling requirements. Suppose there are 4 data point on minority class and 10 data point on majority class. For this imbalance data set, balance by increasing minority class with synthetic data point.   SMOTE creating synthetic data point but it is necessary to consider k nearest neighbours first. If k = 3 then SMOTE consider 3 nearest neighbours. Figure-2 display SMOTE with k = 3 and x = x1, x2, x3, x4 data point denote minority class. And all circles represent majority class.

SMOTE Example

Figure- 2 SMOTE example

 

4 Phase 3: – Feature Reduction

4.1 LDA

LDA stands for Linear Discriminant analysis supervised technique are commonly used for classification problem.  On this feature reduction account continuous independent variable and output categorical variable. It is multivariate analysis technique. LDA analyse by comparing mean of the variables.  Main goal of LDA is differentiate classes in low dimension space. LDA is similar to PCA (Principal component analysis) but in addition LDA maximize the separation between multiple classes. LDA is a dimensionality reduction technique where creating synthetic feature from linear combination of original data set then discard less important feature. LDA calculate class variance, it maximize between class variance and minimize within class variance. Table-2 display the process steps of LDA.

LDA Process

Table- 2 LDA process

5 Phase 5: – Base Model

Here we consider two base model ensemble random forest and MLP classifier.

5.1 Random Forest

Random forest is an ensemble (Bagging) method where group of weak learner (decision tree) come together to form a strong leaner. Random forest is a supervised algorithm which is used for regression and classification problem. Random forests create several decisions tree for predictions and provide solution by voting (classification) or mean (regression) value. Working process of Random forest given below (Table -3).

Random Forest

Table-3 Random Forest process

When training a Random forest root node contains a sample of bootstrap dataset and the feature is as same as original dataset. Suppose the dataset is D and contain d record and m number of columns. From the dataset D random forest first randomly select sample of rows (d) with replacement and sample of features (n) and give it to the decision tree. Suppose Random forest created several decision trees like T1, T2, T3, T4 . . . Tn. Then randomly selected dataset D = d + n is given to the decision tree T1, T2, T3, T4 . . . Tn where D < D, m > n and d > d.  After taking the dataset decision tree give the prediction for binary classification 1 or 0 then aggregating the decision and select the majority voted result. Figure-3 describes the structure of random forest process.

Random Forest Process

Figure- 3 Random Forest process

On Random forest base learner Decision Tree grows complete depth where bias (properly train on training dataset) is low and variance is high (when implementing test data give big error) called overfitting. On Random forest using multiple decision trees where each Decision tree is high variance but when combining all decision trees with the respect of majority vote then high variance converted into low variance because using row and feature sampling with replacement and taking the majority vote where decision is not depend on one decision tree.

CART (Classification and Regression Tree) is binary segmentation technique. CART is a Gini’s impurity index based classical algorithm to split a dataset and build a decision tree. By splitting a selected dataset CART created two child nodes repeatedly and builds a tree until the data no longer be split. There are three steps CART algorithm follow:

  1. Find best split for each features. For each feature in binary split make two groups of the ordered classes. That means possibility of split for k classes is k-1. Find which split is maximized and contain best splits (one for each feature) result.
  2. Find the best split for nodes. From step 1 find the best one split (from all features) which maximized the splitting criterion.
  3. Split the best node from step 2 and repeat from step 1 until fulfil the stopping criterion.

 

For splitting criteria CART use GINI index impurity algorithm to calculate the purity of split in a decision tree. Gini impurity randomly classified the labels with the same distribution in the dataset. A Gini impurity of 0 (lowest) is the best possible impurity and it is achieve when everything is in a same class. Gini index varies from 0 to 1. 0 indicate the purity of class where only one class exits or all element under a specific class. 1 indicates that elements are randomly distributed across various classes. And 0.5 indicate equal elements distributed over classes. Gini index (GI) described by mathematically that sum of squared of probabilities of each class (pi) deducted from one (Equation-5).

Gini Impurities

Equation – 5 Gini impurities

Here (Equation-5) pi represent the probability (probability of p+ or yes and probability of p- or no) of distinct class with classified element. Suppose randomly selected feature (a1) which has 8 yes and 4 no. After the split right had side (b1 on equation-6) has 4 yes and 4 no and left had side (b2 on equation – 7) has 4 yes and 0 no. here b2 is a pure split (leaf node) because only one class yes is present. By using the GI (Gini index) formula for b1 and b2:-

Equation- 6 & 7 – Gini Impurity b1 & Gini Impurity b2

Here for b1 value 0.5 indicates that equal element (yes and no) distribute over classes which is not pure split. And b2 value 0 indicates pure split. On GINI impurity indicates that when probability (yes or no) increases GINI value also increases. Here 0 indicate pure split and .5 indicate equal split that means worst situation. After calculating the GINI index for b1 and b2 now calculate the reduction of impurity for data point a1. Here total yes 8 (b1 and b2 on Equation – 8) and total no 4 (b1) so total data is 12 on a1. Below display the weighted GINI index for feature a1:

Total data point on b1 with Gini index (m) = 8/12 * 0.5 = 0.3333

Total data point on b2 with Gini index (n) = 4/12 * 0 = 0

Weighted Gini index for feature a1 = m + n = 0.3333

Equation- 8 Gini Impurity b1 & b2

After computing the weighted Gini value for every feature on a dataset taking the highest value feature as first node and split accordingly in a decision tree. Gini is less costly to compute.

5.2 Multilayer Perceptron Classifier (MLP Classifier)

Multilayer perceptron classifier is a feedforward neural network utilizes supervised learning technique (backpropagation) for training. MLP Classifier combines with multiple perceptron (hidden) layers. For feedforward taking input send combining with weight bias and then activation function from one hidden layer output goes to other hidden and this process continuing until reached the output. Then output calculates the error with error algorithm. These errors send back with backpropagation for weight adjustment by decreasing the total error and process is repeated, this process is call epoch. Number of epoch is determined with the hyper-parameter and reduction rate of total error.

5.2.1 Back-Propagation

Backpropagation is supervised learning algorithm that is used to train neural network. A neural network consists of input layer, hidden layer and output layer and each layer consists of neuron. So a neural network is a circuit of neurons. Backpropagation is a method to train multilayer neural network the updating of the weights of neural network and is done in such a way so that the error observed can be reduced here, error is only observed in the output layer and that error is back propagated to the previous layers and previous layer is proportionally updated weight. Backpropagation maintain chain rule to update weight. Mainly three steps on backpropagation are (Table-4):

Step Process
Step 1 Forward Pass
Step 2 Backward Pass
Step 3 Sum of all values and calculate updated weight value with Chain – rules.

Table-4 Back-Propagation process

5.2.2 Forward pass/ Forward propagation

Forward propagation is the process where input layer send the input value with randomly selected weight and bias to connected neuron and inside neuron selected activation function combine them and forward to other connected neuron layer after layer then give an output with the help of output layer. Below (Figure-4) display the forward propagation.

Foreward Pass

Figure-4 Forward passes

Input layer take the input of X (X1, X2) combine with randomly selected weight for each connection and with fixed bias (different hidden layer has different bias) send it to first hidden layer where first multiply the input with corresponding weight and added all input with single bias then selected activation function (may different form other layer) combine all input and give output according to function and this process is going on until reach in output layer. Output layer give the output like Y (Y1, Y2) (here output is binary classification as an example) according to selected activation function.

5.2.3 Backward Pass

After calculating error (difference between Forward pass output and actual output) backward pass try to minimize the error with optimisation function by sending backward with proportionally distribution and maintain a chain rule. Backward pass distribution the error in such a way where weighted value is taking under consideration. Below (Figure-5) diagram display the Backward pass process.

Backward Pass

Figure-5 Backward passes

Backpropagation push back the error which is calculated with error function or loss function for update proportional distribution with the help of optimisation algorithm. Division of Optimisation algorithm given below on Figure – 6

Optimisation Algorithms

Figure -6 Division of Optimisation algorithms

Gradient decent calculate gradient and update value by increases or decreases opposite direction of gradients unit and try to find the minimal value. Gradient decent update just one time for whole dataset but stochastic gradient decent update on each training sample and it is faster than normal gradient decent. Gradient decent can be improve by tuning parameter like learning rate (0 to 1 mostly use 0.5). Adagrad use time step based parameter to compute learning rate for every parameter. Adam is Adaptive Moment Estimation. It calculates different parameter with different learning rate. It is faster and performance rate is higher than other optimization algorithm. On the other way Adam algorithm is squares the calculated exponential weighted moving average of gradient.

5.2.4 Chain – rules

Backpropagation maintain chain-rules to update weighted value. On chain-rules backpropagation find the derivative of error respect to any weight. Suppose E is output error. w is weight for input a and bias b and ac neuron output respect of activation function and summation of bias with weighted input (w*a) input to neuron is net. So partial derivative for error respect to weight is ∂E / ∂w display the process on figure-7.

Figure- 7 Partial derivative for error respect to weight

On the chain rules for backward pass to find (error respect to weight) ∂E / ∂w = ∂E / ∂ac * ∂ac / ∂net * ∂net / ∂w. here find to error respect to weight are error respect to output of activation function multiply by activation function output respect to input in a neuron multiply by input in a neuron respect to weight.

5.2.5 Activation function

Activation function is a function which takes the decision about neuron to activate or deactivate. If the activate function activate the neuron then it will give an output on the basis of input. Input in a activation function is sum of input multiply with corresponding weight and adding the layered bias.  The main function of a activate function is non-linearity output of a neuron.

Activation Function

Figure-8 Activation function

Figure – 8 display a neuron in a hidden layer. Here several input (1, 2, 3) with corresponding weight (w1, w2, w3) putting in a neuron input layer where layer bias add with summation of multiplication with input and weight. Equation-9 display the output of an activate function.

Output from activate function y = Activate function (Ʃ (weight * input) + bias)

y = f (Ʃ (w*x) +b)

Equation- 9 Activate function

There are many activation functions like linear function for regression problem, sigmoid function for binary classification problem where result either 0 or 1, Tanh function which is based on sigmoid function but mathematically shifted version and values line -1 to 1. RELU function is Rectified linear unit. RELU is less expensive to compute.

5.2.6 Sigmoid

Sigmoid is a squashing activate function where output range between 0 and 1. Sigmoidal name comes from Greek letter sigma which looks like letter S when graphed. Sigmoid function is a logistic type function, it mainly use in output layer in neural network. Sigmoid is non-linear, fixed output range (between 0 and 1), monotonic (never decrees or never increases) and continuously differentiated function. Sigmoid function is good at classification and output from sigmoid is nonlinear. But Sigmoid has a vanishing gradient problem because output variable is very less to change in input variable. Figure- 9 displays the output of a Sigmoid and derivative of Sigmoid. Here x is any number (positive or negative). On sigmoid function 1 is divided by exponential negative input with adding 1.

Sigmoid

Figure – 9 Sigmoid Functions

4.5.2.7 RELU

RELU stands for Rectified Linear Units it is simple, less expensive in computation and rectifies the gradient vanishing problem. RELU is nonlinear activation function. It gives output either positive (infinity) or 0. RELU has a dying problem because if neurons stop for responding to variation because of gradient is 0 or nothing has to change. Figure- 10 displays the output of an RELU and derivative of RELU. Here x is any positive input and if x is grater then 0 give the output as x or give output 0. RELU function gives the output maximum value of input, here max (0, x).

Relu Activation Function

Figure – 10 RELU Function

4.5.2.8 Cost / loss function (Binary Cross-Entropy)

Cost or loss function compare the predictive value (model outcome) with actual value and give a quantitative value which give the indication about how much good or bad the prediction is.

Cost Function

Figure- 11 Cost function work process

Figure-11 x1 and x2 are input in a activate function f(x) and output y1_out which is sum of weighted input added with bias going through activate function. After model output activate function compare the output with actual output and give a quantitative value which indicate how good or bad the prediction is.

There are many type of loss function but choosing of optimal loss function depends on the problem going to be solved such as regression or classification. For binary classification problem binary cross entropy is used to calculate cost. Equation-10 displays the binary cross entropy where y is actual binary value and yp predictive outcome range 0 and 1. And i is scalar vale range between 1 to model output size (N).

Binary Crossentropy

Equation-10 displays the binary cross entropy

6 Phase 6: – Evaluation

6.1 Confusion matrix

In a classification confusion matrix describe the performance of actual value against predictive value. Confusion Matrix does the performance measurement. So confusion matrix classifies and display predicted and actual value (Visa, S., Ramsay 2011).

Confusion Matrix

Table- 5 Confusion Matrix

Confusion Matrix (Table-5) combines with True Positive (TP), True Negative (TN), False Positive (FP), and False Negative (FN). True Positive is prediction positive and true. True Negative is prediction negative and that is true. False positive is prediction positive and it’s false. False negative is prediction negative and that is false. False positive is known as Type1 error and false negative is known as Type 2 error. Confusion matrix can able to calculate several list of rates which are given below on Table- 6.

Here    N = Total number of observation, TP = True Positive, TN = True Negative

FP = False Positive, FN = False Negative, Total Actual No (AN) = TN + FP,

Total Predictive Yes (PY) = FP + TP. Total Actual Yes (AY) = FN + TP

Rate

 

Description Mathematical Description
Accuracy Classifier, overall how often correctly identified  (TP+TN) / N
Misclassification Rate Classifier, overall how often wrongly identified (FP + FN) / N
True Positive Rate

(Sensitivity / Recall)

Classifier, how often predict correctly yes when it is actually yes.  TP / AY
False Positive Rate Classifier, how often predict wrongly yes when it is actually no.  FP / AN
True Negative Rate

(Specificity)

Classifier, how often predict correctly no when it is actually no.  TN / AN
Precision Classifier how often predict yes when it is correct.  TP / PY
Prevalence Yes conditions how often occur in a sample. AY / N

Table – 6 Confusion matrixes Calculation

From confusion matrix F1 score can be calculated because F1 score related to precision and recall. Higher F1 score is better. If precision or recall any one goes down F1 score also go down.

F1 = \frac{2 * Precision * Recall}{Precision + Recall}

4.6.2 ROC (Receiver Operating Characteristic) curve

In statistics ROC is represent in a graph with plotting a curve which describe a binary classifiers performance as its differentiation threshold is varied. ROC (Equation-11) curve created true positive rate (TPR) against false positive rate (FPR). True positive rate also called as Sensitivity and False positive rate also known as Probability of false alarm. False positive rate also called as a probability of false alarm and it is calculated as 1 – Specificity.

True Positive Rate = \frac{True Positive}{True Positive + False Negative} = Recall or Sensivity

False Positive Rate = \frac{True Negative}{True Negative + False Positive} = 1 - Specificity

Equation- 11 ROC

So ROC (Receiver Operation Characteristic) curve allows visual representation between sensitivity and specificity associated with different values of the test result (Grzybowski, M. and Younger, J.G., 1997)

On ROC curve each point has different Threshold level. Below (Figure – 12) display the ROC curve. Higher the area curve covers is better that means high sensitivity and high specificity represent more accuracy. ROC curve also represent that if classifier predict more often true than it has more true positive and also more false positive. If classifier predict true less often then fewer false positive and also fewer true positive.

ROC Curve

ROC Curve

Figure – 12 ROC curve description

4.6.3 AUC (Area under Curve)

Area under curve (AUC) is the area surrounded by the ROC curve and AUC also represent the degree of separability that means how good the model to distinguished between classes. Higher the AUC value represents better the model performance to separate classes. AUC = 1 for perfect classifier, AUC = 0 represent worst classifier, and AUC = 0.5 means has no class separation capacity. Suppose AUC value is 0.6 that means 60% chance that model can classify positive and negative class.

Figure- 13 to Figure – 16 displays an example of AUC where green distribution curve for positive class and blue distribution curve for negative class. Here threshold or cut-off value is 0.5 and range between ‘0’ to ‘1’. True negative = TN, True Positive = TP, False Negative = FN, False Positive = FP, True positive rate = TPR (range 0 to 1), False positive rate = FPR (range 0 to 1).

On Figure – 13 left distribution curve where two class curves does not overlap that means both class are perfectly distinguished. So this is ideal position and AUC value is 1.  On the left side ROC also display that TPR for positive class is 100% occupied.

ROC distributions (perfectly distinguished

ROC distributions (perfectly distinguished

Figure – 14 two class overlap each other and raise false positive (Type 1), false negative (Type 2) errors. Here error could be minimize or maximize according to threshold. Suppose here AUC = 0.6, that means chance of a model to distinguish two classes is 60%. On ROC curve also display the curve occupied for positive class is 60%.

ROC distributions (class partly overlap distinguished)

ROC distributions (class partly overlap distinguished)

Figure- 15 displayed that positive and negative overlap each other. Here AUC value is 0.5 or near to 0.5. On this position classifier model does not able distinguish positive and negative classes. On left side ROC curve become straight that means TPR and FPR are equal.

ROC distributions (class fully overlap distinguished)

ROC distributions (class fully overlap distinguished)

Figure- 16 positive and negative class swap position and on this position AUC = 0. That means classified model predict positive as a negative and negative as a positive. On the left ROC curve display that curve on FPR side fully fitted.

ROC distributions (class swap position distinguished)

ROC distributions (class swap position distinguished)

7 Summaries

This paper describes a data mining process flow and related model and its algorithm with textual representation. One hot encoding create dummy variable for class features and min-max scaling scale the data in a single format. Balancing by SMOTE data where Euclidian distance calculates the distance in-between nearest neighbour to produce synthetic data under minority class. LDA reduce the distance inside class and maximise distance in-between class and for two class problem give a single dimension features which is less costly to calculate accuracy by base algorithm (random forest and MLP classifier).  Confusion matrix gives the accuracy, precision, sensitivity, specificity which is help to take a decision about base algorithm. AUC and ROC curve also represent true positive rate against false positive rate which indicate base algorithm performance.

Base algorithm Random forest using CART with GINI impurity for feature selection to spread the tree. Here CART is selected because of less costly to run. Random forest algorithm is using bootstrap dataset to grow trees, and aggregation using majority vote to select accuracy.

MLP classifier is a neural network algorithm using backpropagation chain-rule to reducing error. Here inside layers using RLU activation function. Output layers using Sigmoid activation function and binary cross entropy loss function calculate the loss which is back propagate with Adam optimizer to optimize weight and reduce loss.

References:

  1. Visa, S., Ramsay, B., Ralescu, A.L. and Van Der Knaap, E., 2011. Confusion Matrix-based Feature Selection. MAICS, 710, pp.120-127.
  2. Grzybowski, M. and Younger, J.G., 1997. Statistical methodology: III. Receiver operating characteristic (ROC) curves. Academic Emergency Medicine, 4(8), pp.818-826.

CRISP-DM methodology in technical view

On this paper discuss about CRISP-DM (Cross Industry Standard Process for data mining) methodology and its steps including selecting technique to successful the data mining process. Before going to CRISP-DM it is better to understand what data mining is? So, here first I introduce the data mining and then discuss about CRISP-DM and its steps for any beginner (data scientist) need to know.

1 Data Mining

Data mining is an exploratory analysis where has no idea about interesting outcome (Kantardzic, 2003). So data mining is a process to explore by analysis a large set of data to discover meaningful information which help the business to take a proper decision. For better business decision data mining is a way to select feature, correlation, and interesting patterns from large dataset (Fu, 1997; SPSS White Paper, 1999).

Data mining is a step by step process to discover knowledge from data. Pre-processing data is vital part for a data mining. In pre-process remove noisy data, combining multiple sources of data, retrieve relevant feature and transforming data for analysis. After pre-process mining algorithm applied to extract data pattern so data mining is a step by step process and applied algorithm to find meaning full data pattern. Actually data mining is not only conventional analysis it is more than that (Read, 1999).

Data mining and statistics closely related. Main goal of data mining and statistic is find the structure of data because data mining is a part of statistics (Hand, 1999). However, data mining use tools, techniques, database, machine learning which not part of statistics but data mining use statistics algorithm to find a pattern or discover hidden decision.

Data mining objective could be prediction or description. On prediction data mining considering several features of dataset to predict unidentified future, on the other hand description involve identifying pattern of data to interpreted (Kantardzic, 2003).

From figure 1.1 shows data mining is the only one part of getting unknown information from data but it is the central process of whole process. Before data mining there are several processes need to be done like collecting data from several sources than integrated data and keep in data storage. Stored unprocessed data evaluated and selected with pre-processed activity to give a standard format than data mining algorithm to analysis for hidden pattern.

Data Mining Process

2 CRISP-DM Methodologies

Cross Industry Standard Process for data mining (CRISP-DM) is most popular and widely uses data mining methodology. CRISP-DM breaks down the data mining project life cycle into six phases and each phase consists of many second-level generic tasks. Generic task cover all possible data mining application. CRISP-DM extends KDD (Knowledge Discovery and Data Mining) into six steps which are sequence of data mining application (Martínez-Plumed 2019).

Data science and data mining project extract meaningful information from data. Data science is an art where a lot of time need to spend for understanding the business value and data before applying any algorithm then evaluate and deployed a project. CRISP-DM help any data science and data mining project from start to end by giving step by step process.

Present world every day billions of data are generating. So organisations are struggling with overwhelmed data to process and find a business goal. Comprehensive data mining methodology, CRISP-DM help business to achieve desirable goal by analysing data.

CRISP-DM (Cross Industry Standard Process for Data Mining) is well documented, freely available, data mining methodology. CRISP-DM is developed by more than 200 data mining users and many mining tool and service providers funded by European Union. CRISP-DM encourages organization for best practice and provides a structure of data mining to get better, faster result.

CRISP-DM is a step by step methodology. Figure-2.1 show the phases of CRISP-DM and process of data mining. Here one side arrow indicates the dependency between phases and double side arrow represents repeatable process. Six phases of CRISP-DM are Business understanding, Data understanding, Modelling, Evaluation and Deployment.

CRISP-DM

2.1 Business Understanding

Business Understanding or domain understanding is the first step of CRISP-DM methodology. On this stage identify the area of business which is going to transform into meaningful information by analysing, processing and implementing several algorithms. Business understanding identifies the available resource (human and hardware), problems and set a goal. Identification of business objective should be agreed with project sponsors and other unit of business which will be affected. This step also focuses about details business success criteria, requirements, constraints, risk, project plan and timeline.

2.2 Data Understanding

Data understanding is the second and closely related with the business understanding phase. This phase mainly focus on data collection and proceeds to get familiar with the data and also detect interesting subset from data. Data understanding has four subsets these are:-

2.2.1 Initial data collection

On this subset considering the data collection sources which is mainly divided into two categories like outsource data or internal source data.  If data is from outsource then it may costly, time consuming and may be low quality but if data is collected form internal source it is an easy and less costly, but it may be contain irrelevant data. If internal source data does not fulfil the interest of analysis than it is necessary to move outsource data. Data collection also give an assumption that the data is quantitative (continuous, count) or qualitative (categorical).  It also gives information about balance or imbalanced dataset.  On data collection should avoid random error, systematic error, exclusion errors, and errors of choosing.

2.2.2 Data Description

Data description performs initial analysis about data. On this stage it is going to determine about the source of data like RDBMS, SQL, NoSQL, Big data etc. then analysis and describe the data about size (large data set give more accurate result but time consuming), number of records, tables, database, variables, and data types (numeric, categorical or Boolean). On this phase examine the accessibility and availability of attributes.

2.2.3 Exploratory data analysis (EDA)

On exploratory data analysis describe the inferential statistics, descriptive statistics and graphical representation of data. Inferential statistics summarize the entire population from the sample data to perform sampling and hypothesis testing. On Parametric hypothesis testing  (Null or alternate – ANOVA, t-test, chi square test) perform for known distribution (based on population) like mean, variance, standard deviation, proportion and Non-parametric hypothesis testing perform when distribution is unknown or sample size is small. On sample dataset, random sampling implement when dataset is balance but for imbalance dataset should be follow random resampling (under  and over sampling), k fold cross validation, SMOTE (synthetic minority oversampling technique), cluster base sampling, ensemble techniques (bagging and boosting – Add boost, Gradient Tree Boosting, XG Boost) to form a balance dataset.

On descriptive statistics analysis describe about the mean, median, mode for measures of central tendency on first moment business decision. On second moment business decision describe the measure of dispersion about the variance, standard deviation and range of data.  On third and fourth moment business decision describe accordingly skewness (Positive skewness – heavier tail to the right, negative skewness – heavier tail to the left, Zero skewness – symmetric distribution) and Kurtosis (Leptokurtosis – heavy tail, platykurtosis – light tail, mesokurtic – normal distribution).

Graphical representation is divided into univariate, bivariate and multivariate analysis. Under univariate whisker plot, histogram identify the outliers and shape of distribution of data and Q-Q plot (Quantile – Quantile) plot describe the normality of data that means data is normally distribution or not.  On whisker plot if data present above of Q3 + 1.5 (IQR) and below of Q1 – 1.5 (IQR) is outlier. For Bivariate correlations identify with scatter plot which describe positive, negative or no correlation and also identify the data linearity or non-linearity. Scatter plot also describe the clusters and outliers of data.  For multivariate has no graphical analysis but used to use regression analysis, ANOVA, Hypothesis analysis.

2.2.4 Data Quality analysis

This phase identified and describes the potential errors like outliers, missing data, level of granularity, validation, reliability, bad metadata and inconsistency.  On this phase AAA (attribute agreement analysis) analysed discrete data for data error. Continuous data analysed with Gage repeatability and reproducibility (Gage R & R) which follow SOP (standard operating procedures). Here Gage R & R define the aggregation of variation in the measurement data because of the measurement system.

2.3 Data Preparation

Data Preparation is the time consuming stage for every data science project. Overall on every data science project 60% to 70% time spend on data preparation stage. Data preparation stapes are described below.

2.3.1 Data integration

Data integration involved to integrate or merged multiple dataset. Integration integrates data from different dataset where same attribute or same columns presents but when there is different attribute then merging the both dataset.

2.3.2 Data Wrangling

On this subset data are going to clean, curate and prepare for next level. Here analysis the outlier and treatment done with 3 R technique (Rectify, Remove, Retain) and for special cases if there are lots of outliner then need to treat outlier separately (upper outliner in an one dataset and lower outliner in another dataset) and alpha (significant value) trim technique use to separate the outliner from the original dataset. If dataset has a missing data then need to use imputation technique like mean, median, mode, regression, KNN etc.

If dataset is not normal or has a collinearity problem or autocorrelation then need to implement transformation techniques like log, exponential, sort, Reciprocal, Box-cox etc. On this subset use the data normalization (data –means/standard deviation) or standardization (min- max scaler) technique to make unitless and scale free data. This step also help if data required converting into categorical then need to use discretization or binning or grouping technique. For factor variable (where has limited set of values), dummy variable creation technique need to apply like one hot encoding.  On this subset also help heterogeneous data to transform into homogenous with clustering technique. Data inconsistencies also handle the inconsistence of data to make data in a single scale.

2.3.3 Feature engineering and selection/reduction

Feature engineering may called as attribute generation or feature extraction. Feature extraction creating new feature by reducing original feature to make simplex model. Feature engineering also do the normalized feature by producing calculative new feature. So feature engineering is a data pre-process technique where improve data quality by cleaning, integration, reduction, transformation and scaling.

Feature selections reduce the multicollinearity or high correlated data and make model simple. Main two type of feature selection technique are supervised and unsupervised. Principal Components Analysis (PCA) is an unsupervised feature reduction/ feature selection technique and LDA is a Linear Discriminant analysis supervised technique mainly use for classification problem. LDA analyse by comparing mean of the variables. Supervised technique is three types filter, wrapper and ensemble method. Filter method is easy to implement but wrapper is costly method and ensemble use inside a model.

2.4 Model

2.4.1 Model Selection Technique

Model selection techniques are influence by accuracy and performance.  Because recommendation need better performance but banking fraud detection needs better accuracy technique.  Model is mainly subdivided into two category supervised learning where predict an output variable according to given an input variable and unsupervised learning where has not output variable.

On supervised learning if an output variable is categorical than it is classification problem like two classes or multiclass classification problem. If an output variable is continuous (numerical) then the problem is called prediction problem. If need to recommending according to relevant information is called recommendation problem or if need to retrieve data according to relevance data is called retrieval problem.

On unsupervised learning where target or output variable is not present. On this technique all variable is treated as an input variable. Unsupervised learning also called clustering problem where clustering the dataset for future decision.

Reinforcement learning agent solves the problem by getting reward for success and penalty for any failure. And semi-supervised learning is a process to solve the problem by combining supervised and unsupervised learning method. On semi-supervised, a problem solved by apply unsupervised clustering technique then for each cluster apply different type of supervised machine learning algorithm like linear algorithm, neural network, K nearest  neighbour etc.

On data mining model selection technique, where output variable is known, then need to implement supervised learning.  Regression is the first choice where interpretation of parameter is important. If response variable is continuous then linear regression or if response variable is discrete with 2 categories value then logistic regression or if response variable is discrete with more than 2 categorical values then multinomial or ordinal regression or if response variable is count then poission where mean is equal to variance or negative binomial regression where variance is grater then mean or if response variable contain excessive zero values then need to choose Zero inflated poission (ZIP) or Zero inflated negative binomial (ZINB).

On supervised technique except regression technique all other technique can be used for both continuous or categorical response variable like KNN (K-Nearest Neighbour),  Naïve Bays, Black box techniques (Neural network, Support vector machine), Ensemble Techniques (Stacking, Bagging like random forest, Boosting like Decision tree, Gradient boosting, XGB, Adaboost).

When response variable is unknown then need to implement unsupervised learning. Unsupervised learning for row reduction is K-Means, Hierarchical etc., for columns reduction or dimension reduction PCA (principal component analysis), LDA (Linear Discriminant analysis), SVD (singular value decomposition) etc. On market basket analysis or association rules where measure are support and confidence then lift ration to determine which rules is important. There are recommendation systems, text analysis and NLP (Natural language processing) also unsupervised learning technique.

For time series need to select forecasting technique. Where forecasting may model based or data based. For Trend under model based need to use linear, exponential, quadratic techniques. And for seasonality need to use additive, multiplicative techniques. On data base approaches used auto regressive, moving average, last sample, exponential smoothing (e.g. SES – simple exponential smoothing, double exponential smoothing, and winters method).

2.4.2 Model building

After selection model according to model criterion model is need to be build. On model building provided data is subdivided with training, validation and testing.  But sometime data is subdivided just training and testing where information may leak from testing data to training data and cause an overfitting problem. So training dataset should be divided into training and validation whereas training model is tested with validation data and if need any tuning to do according to feedback from validation dataset. If accuracy is acceptable and error is reasonable then combine the training and validation data and build the model and test it on unknown testing dataset. If the training error and testing error is minimal or reasonable then the model is right fit or if the training error is low and testing error is high then model is over fitted (Variance) or if training error is high and testing error is also high then model is under fitted (bias). When model is over fitted then need to implement regularization technique (e.g. linear – lasso, ridge regression, Decision tree – pre-pruning, post-pruning, Knn – K value, Naïve Bays – Laplace, Neural network – dropout, drop connect, batch normalization, SVM –  kernel trick)

When data is balance then split the data training, validation and testing and here training is larger dataset then validation and testing. If data set is imbalance then need to use random resampling (over and under) by artificially increases training dataset. On random resampling by randomly partitioning data and for each partition implement the model and taking the average of accuracy. Under K fold cross validation creating K times cross dataset and creating model for every dataset and validate, after validation taking the average of accuracy of all model. There is more technique for imbalance dataset like SMOTH (synthetic minority oversampling technique), cluster based sampling, ensemble techniques e.g. Bagging, Boosting (Ada Boost, XGBoost).

2.4.3 Model evaluation and Tuning

On this stage model evaluate according to errors and accuracy and tune the error and accuracy for acceptable manner. For continuous outcome variable there are several way to measure the error like mean error, mean absolute deviation, Mean squared error, Root mean squared error, Mean percentage error and Mean absolute percentage error but more acceptable way is Mean absolute percentage error. For this continuous data if error is known then it is easy to find out the accuracy because accuracy and error combining value is one. The error function also called cost function or loss function.

For discrete output variable model, for evaluation and tuning need to use confusion matrix or cross table. From confusion matrix, by measuring accuracy, error, precision, sensitivity, specificity, F1 help to take decision about model fitness. ROC curve (Receiver operating characteristic curve), AUC curve (Area under the ROC curve) also evaluate the discrete output variable. AUC and ROC curve plot of sensitivity (true positive rate) vs 1-specificity (false positive rate).  Here sensitivity is a positive recall and  recall is basically out of all positive samples, how sample classifier able to identify. Specificity is negative recall here recall is out of all negative samples, how many sample classifier able to identify.  On AUC where more the area under the ROC is represent better accuracy. On ROC were step bend it’s indicate the cut off value.

2.4.4 Model Assessment

There is several ways to assess the model. First it is need to verify model performance and success according to desire achievement. It needs to identify the implemented model result according to accuracy where accuracy is repeatable and reproducible. It is also need to identify that the model is scalable, maintainable, robust and easy to deploy. On assessment identify that the model evaluation about satisfactory results (identify the precision, recall, sensitivity are balance) and meet business requirements.

2.5 Evaluation

On evaluation steps, all models which are built with same dataset, given a rank to find out the best model by assessing model quality of result and simplicity of algorithm and also cost of deployment. Evaluation part contains the data sufficiency report according to model result and also contain suggestion, feedback and recommendation from solutions team and SMEs (Subject matter experts) and record all these under OPA (organizational process assets).

2.6 Deployment

Deployment process needs to monitor under PEST (political economical social technological) changes within the organization and outside of the organization. PEST is similar to SWOT (strength weakness opportunity and thread) where SW represents the changes of internal and OT represents external changes.

On this deployment steps model should be seamless (like same environment, same result etc.) from development to production. Deployment plan contain the details of human resources, hardware, software requirements. Deployment plan also contain maintenance and monitoring plan by checking the model result and validity and if required then implement retire, replace and update plan.

3 Summaries

CRISP-DM implementation is costly and time consuming. But CRISP-DM methodology is an umbrella for data mining process. CRISP-DM has six phases, Business understanding, Data understanding, Modelling, Evaluation and Deployment. Every phase has several individual criteria, standard and process. CRISP-DM is Guideline for data mining process so if CRISP-DM is going to implement in any project it is necessary to follow each and every single guideline and maintain standard and criteria to get required result.

4 References

  1. Fu, Y., (1997), “Data Mining: Tasks, Techniques and Applications”, Potentials, IEEE, 16: 4, 18–20.
  2. Hand, D. J., (1999), “Statistics and Data Mining: Intersecting Disciplines”, ACM SIGKDD Explorations Newsletter, 1: 1, 16 – 19.
  3. Kantardzic, M., (2003), “Data Mining: Concepts, Models, Methods, and Algorithms” John Wiley and Sons, Inc., Hoboken, New Jersey
  4. Martínez-Plumed, F., Contreras-Ochando, L., Ferri, C., Orallo, J.H., Kull, M., Lachiche, N., Quintana, M.J.R. and Flach, P.A., 2019. CRISP-DM Twenty Years Later: From Data Mining Processes to Data Science Trajectories. IEEE Transactions on Knowledge and Data Engineering.
  5. Read, B.J., (1999), “Data Mining and Science? Knowledge discovery in science as opposed to business”, 12th ERCIM Workshop on Database Research.

The algorithm known as PCA and my taxonomy of linear dimension reductions

In one of my previous articles, I explained the importance of reducing dimensions. Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA) are the simplest types of dimension reduction algorithms. In upcoming articles of mine, you are going to see what these algorithms do. In conclusion, diagonalization, which I mentioned in the last article, is what these algorithms are all about, but still in this article I can mainly cover only PCA.

This article is largely based on the explanations in Pattern Recognition and Machine Learning by C. M. Bishop (which is often called “PRML”), and when you search “PCA” on the Internet, you will find more or less similar explanations. However I hope I can go some steps ahead throughout this article series. I mean, I am planning to also cover more generalized versions of PCA, meanings of diagonalization, the idea of subspace. I believe this article series is also effective for refreshing your insight into linear algebra.

*This is the third article of my article series “Illustrative introductions on dimension reduction.”

1. My taxonomy on linear dimension reduction

*If you soon want to know  what the algorithm called “PCA” is, you should skip this section for now to avoid confusion.

Out of the two algorithms I mentioned, PCA is especially important and you would see the same or similar ideas in various fields such as signal processing, psychology, and structural mechanics. However in most cases, the word “PCA” refers to one certain algorithm of linear dimension reduction. Most articles or study materials only mention the “PCA,” and this article is also going to cover only the algorithm which most poeple call “PCA.” However I found that PCA is only one branch of linear dimension reduction algorithms.

*From now on all the terms “PCA” in this article means the algorithm known as PCA unless I clearly mention the generalized KL transform.

*This chart might be confusing to you. According to PRML, PCA and KL transform is identical. PCA has two formulations, maximum variance formulation and minimum error formulation, and they can give the same result. However according to a Japanese textbook, which is very precise about this topic, KL transform has two formulations, and what we call PCA is based on maximum variance formulation. I am still not sure about correct terminology, but in this article I am going to call the most general algorithm “generalized KL transform,” I mean the root of the chart above.

*Most materials just explain the most major PCA, but if you consider this generalized KL transform, I can introduce an intriguing classification algorithm called subspace method. This algorithm was invented in Japan, and this is not so popular in machine learning textbooks in general, but learning this method would give you better insight into the idea of multidimensional space in machine learning. In the future, I am planning to cover this topic in this article series.

2. PCA

When someones mention “PCA,” I am sure for the most part that means the algorithm I am going to explain in the rest of this article. The most intuitive and straightforward way to explain PCA is that, PCA (Principal Component Analysis) of two or three dimensional data is fitting an oval to two dimensional data or fitting an ellipsoid to three dimensional data. You can actually try to plot some random dots on a piece of paper, and draw an oval which fits the dots the best. Assume that you have these 2 or 3 dimensional data below, and please try to put an oval or an ellipsoid to the data.

I think this is nothing difficult, but I have a question: what was the logic behind your choice?

Some might have roughly drawn its outline. Formulas of  “the surface” of general ellipsoids can be explained in several ways, but in this article you only have to consider ellipsoids whose center is the origin point of the coordinate system. In PCA you virtually shift data so that the mean of the data comes to the origin point of the coordinate system. When A is a certain type of D\times D matrix, the formula of a D-dimensional ellipsoid whose center is identical to the origin point is as follows: (\boldsymbol{x}, A\boldsymbol{x}) = 1, where \boldsymbol{x}\in \mathbb{R}. As is always the case with formulas in data science, you can visualize such ellipsoids if you are talking about 1, 2, or 3 dimensional data like in the figure below, but in general D-dimensional space, it is theoretical/imaginary stuff on blackboards.

*In order to explain the conditions which the matrix A has to hold, I need another article, so for now please just assume that the A is a kind of magical matrix.

You might have seen equations of 2 or 3 dimensional ellipsoids in the following way: \frac{x^2}{a^2} + \frac{y^2}{b^2} = 1, where a\neq 0, b\neq 0 or \frac{x^2}{a^2} + \frac{y^2}{b^2} + \frac{z^2}{c^2}= 1, where a\neq 0, b\neq 0, c \neq 0. These are special cases of the equation (\boldsymbol{x}, A\boldsymbol{x}) = 1, where A=diag(a_1^2, \dots, a_D^2). In this case the axes of ellipsoids the same as those of the coordinate system. Thus in this simple case, A=diag(a^2, b^2) or A=diag(a^2,c^2,c^2).

I am going explain these equations in detail in the upcoming articles. But thre is one problem: how would you fit an ellipsoid when a data distribution does not look like an ellipsoid?

In fact we have to focus more on another feature of ellipsoids: all the axes of an ellipsoid are orthogonal. In conclusion the axes of the ellipsoids are more important in PCA, so I do want you to forget about the surface of ellipsoids for the time being. You might get confused if you also think about the surface of ellipsoid now. I am planning to cover this topic in the next article. I hope this article, combined with the last one and the next one, would help you have better insight into the ideas which frequently appear in data science or machine learning context.

3. Fitting orthogonal axes on data

*If you have no trouble reading the chapter 12.1 of PRML, you do not need to this section or maybe even this article, but I hope at least some charts or codes of mine would enhance your understanding on this topic.

*I must admit I wrote only the essence of PCA formulations. If that seems too abstract to you, you should just breifly read through this section and go to the next section with a more concrete example. If you are confused, there should be other good explanations on PCA on the internet, and you should also check them. But at least the visualization of PCA in the next section would be helpful.

As I implied above, all the axes of ellipsoids are orthogonal, and selecting the orthogonal axes which match data is what PCA is all about. And when you choose those orthogonal axes, it is ideal if the data look like an ellipsoid. Simply putting we want the data to “swell” along the axes.

Then let’s see how to let them “swell,” more mathematically. Assume that you have 2 dimensional data plotted on a coordinate system (\boldsymbol{e}_1, \boldsymbol{e}_2) as below (The samples are plotted in purple). Intuitively, the data “swell” the most along the vector \boldsymbol{u}_1. Also  it is clear that \boldsymbol{u}_2 is the only vector orthogonal to \boldsymbol{u}_1. We can expect that the new coordinate system (\boldsymbol{u}_1, \boldsymbol{u}_2) expresses the data in a better way, and you you can get new coordinate points of the samples by projecting them on new axes as done with yellow lines below.

Next, let’s think about a case in 3 dimensional data. When you have 3 dimensional data in a coordinate system (\boldsymbol{e}_1, \boldsymbol{e}_2,\boldsymbol{e}_3) as below,  the data “swell” the most also along \boldsymbol{u}_1. And the data swells the second most along \boldsymbol{u}_2. The two axes, or vectors span the plain in purple. If you project all the samples on the plain, you will get 2 dimensional data at the right side. It is important that we did not consider the third axis. That implies you might be able to display the data well with only 2 dimensional sapce, which is spanned by the two axes \boldsymbol{v}_1, \boldsymbol{v}_2.

 

Thus the problem is how to calculate such axis \boldsymbol{u}_1. We want the variance of data projected on \boldsymbol{u}_1 to be the biggest. The coordinate of \boldsymbol{x}_n on the axis \boldsymbol{u}_1. The coordinate of a data point \boldsymbol{x}_n on the axis \boldsymbol{u}_1 is calculated by projecting \boldsymbol{x}_n on \boldsymbol{u}_1. In data science context, such projection is synonym to taking an inner  product of \boldsymbol{x}_n and \boldsymbol{u}_1, that is calculating \boldsymbol{u}_1^T \boldsymbol{x}_n.

*Each element of \boldsymbol{x}_n is the coordinate of the data point \boldsymbol{x}_n in the original coordinate system. And the projected data on \boldsymbol{u}_1 whose coordinates are 1-dimensional correspond to only one element of transformed data.

To calculate the variance of projected data on \boldsymbol{u}_1, we just have to calculate the mean of variances of 1-dimensional data projected on \boldsymbol{u}_1. Assume that \bar{\boldsymbol{x}} is the mean of data in the original coordinate, then the deviation of \boldsymbol{x}_1 on the axis \boldsymbol{u}_1 is calculated as \boldsymbol{u}_1^T \boldsymbol{x}_n - \boldsymbol{u}_1^T \bar{\boldsymbol{x}}, as shown in the figure. Hence the variance, I mean the mean of the deviation on is \frac{1}{N} \sum^{N}_{n}{\boldsymbol{u}_1^T \boldsymbol{x}_n - \boldsymbol{u}_1^T \bar{\boldsymbol{x}}}, where N is the total number of data points. After some deformations, you get the next equation \frac{1}{N} \sum^{N}_{n}{\boldsymbol{u}_1^T \boldsymbol{x}_n - \boldsymbol{u}_1^T \bar{\boldsymbol{x}}} = \boldsymbol{u}_1^T S \boldsymbol{u}_1, where S = \frac{1}{N}\sum_{n=1}^{N}{(\boldsymbol{x}_n - \bar{\boldsymbol{x}})(\boldsymbol{x}_n - \bar{\boldsymbol{x}})^T}. S is known as a covariance matrix.

We are now interested in maximizing the variance of projected data on  \boldsymbol{u}_1^T S \boldsymbol{u}_1, and for mathematical derivation we need some college level calculus, so if that is too much for you, you can skip reading this part till the next section.

We now want to calculate \boldsymbol{u}_1 with which \boldsymbol{u}_1^T S \boldsymbol{u}_1 is its maximum value. General \boldsymbol{u}_i including \boldsymbol{u}_1 are just coordinate axes after PCA, so we are just interested in their directions. Thus we can set one constraint \boldsymbol{u}_1^T  \boldsymbol{u}_1 = 1. Introducing a Lagrange multiplier, we have only to optimize next problem: \boldsymbol{u}_1 ^ {*} = \mathop{\rm arg~max}\limits_{\boldsymbol{u}_1} \{ \boldsymbol{u}_1^T S \boldsymbol{u}_1 + \lambda_1 (1 - \boldsymbol{u}_1^T \boldsymbol{u}_1) \}. In conclusion \boldsymbol{u}_1 ^ {*} satisfies S\boldsymbol{u}_1 ^ {*}  = \lamba_1 \boldsymbol{u}_1 ^ {*}. If you have read my last article on eigenvectors, you wold soon realize that this is an equation for calculating eigenvectors, and that means \boldsymbol{u}_1 ^ {*} is one of eigenvectors of the covariance matrix S. Given the equation of eigenvector the next equation holds \boldsymbol{u}_1 ^ {*}^T S \boldsymbol{u}_1 ^ {*} = \lambda_1. We have seen that \boldsymbol{u}_1 ^T S \boldsymbol{u}_1 ^ is a the variance of data when projected on a vector \boldsymbol{u}_1, thus the eigenvalue \lambda_1 is the biggest variance possible when the data are projected on a vector.

Just in the same way you can calculate the next biggest eigenvalue \lambda_2, and it it the second biggest variance possible, and in this case the date are projected on \boldsymbol{u}_2, which is orthogonal to \boldsymbol{u}_1. As well you can calculate orthogonal 3rd 4th …. Dth eigenvectors.

*To be exact I have to explain the cases where we can get such D orthogonal eigenvectors, but that is going to be long. I hope I can to that in the next article.

4. Practical three dimensional example of PCA

We have seen that PCA is sequentially choosing orthogonal axes along which data points swell the most. Also we have seen that it is equal to calculating eigenvalues of the covariance matrix of the data from the largest to smallest one. From now on let’s work on a practical example of data. Assume that we have 30 students’ scores of Japanese, math, and English tests as below.

* I think the subject “Japanese” is equivalent to “English” or “language art” in English speaking countries, and maybe “Deutsch” in Germany. This example and the explanation are largely based on a Japanese textbook named 「これなら分かる応用数学教室 最小二乗法からウェーブレットまで」. This is a famous textbook with cool and precise explanations on mathematics for engineering. Partly sharing this is one of purposes of this article.

At the right side of the figure below is plots of the scores with all the combinations of coordinate axes. In total 9 inverse graphs are symmetrically arranged in the figure, and it is easy to see that English & Japanese or English and math have relatively high correlation. The more two axes have linear correlations, the bigger the covariance between them is.

In the last article, I visualized the eigenvectors of a 3\times 3 matrix A = \frac{1}{50} \begin{pmatrix} 60.45 &  33.63 & 46.29 \\33.63 & 68.49 & 50.93 \\ 46.29 & 50.93 & 53.61 \end{pmatrix}, and in fact the matrix is just a constant multiplication of this covariance matrix. I think now you understand that PCA is calculating the orthogonal eigenvectors of covariance matrix of data, that is diagonalizing covariance matrix with orthonormal eigenvectors. Hence we can guess that covariance matrix enables a type of linear transformation of rotation and expansion and contraction of vectors. And data points swell along eigenvectors of such matrix.

Then why PCA is useful? In order to see that at first, for simplicity assume that x, y, z denote Japanese, Math, English scores respectively. The mean of the data is \left( \begin{array}{c} \bar{x} \\ \bar{y} \\ \bar{z} \end{array} \right) = \left( \begin{array}{c} 58.1 \\ 61.8 \\ 67.3 \end{array} \right), and the covariance matrix of data in the original coordinate system is V_{xyz} = \begin{pmatrix} 60.45 & 33.63 & 46.29 \\33.63 & 68.49 & 50.93 \\ 46.29 & 50.93 & 53.61 \end{pmatrix}. The eigenvalues of  V_{xyz} are \lambda_1=148.34, \lambda_2 = 30.62, and \lambda_3 = 3.60, and their corresponding unit eigenvectors are \boldsymbol{u}_1 =  \left( \begin{array}{c} 0.540 \\ 0.602 \\ 0.589 \end{array} \right) , \boldsymbol{u}_2 =  \left( \begin{array}{c} 0.736 \\ -0.677 \\ 0.0174 \end{array} \right) , \boldsymbol{u}_3 =  \left( \begin{array}{c} -0.408 \\ -0.4.23 \\ 0.809 \end{array} \right) respectively.  U = (\boldsymbol{u}_1 \quad \boldsymbol{u}_2 \quad \boldsymbol{u}_3 )  is an orthonormal matrix, where \boldsymbol{u}_i^T\boldsymbol{u}_j = \begin{cases} 1 & (i=j) \\ 0 & (otherwise) \end{cases}. As I explained in the last article, you can diagonalize V_{xyz} with U: U^T V_{xyz}U = diag(\lambda_1, \dots, \lambda_D).

In order to see how PCA is useful, assume that \left( \begin{array}{c} \xi \\ \eta \\ \zeta \end{array} \right)  = U^T \left( \begin{array}{c} x - \bar{x} \\ y - \bar{y} \\ z - \bar{z} \end{array} \right).

Let’s take a brief look at what a linear transformation by U^T means. Each element of \boldsymbol{x} denotes coordinate of the data point \boldsymbol{x}  in the original coordinate system (In this case the original coordinate system is composed of \boldsymbol{e}_1, \boldsymbol{e}_2, and \boldsymbol{e}_3). U = (\boldsymbol{u}_1, \boldsymbol{u}_2, \boldsymbol{u}_3) enables a rotation of a rigid body, which means the shape or arrangement of data will not change after the rotation, and U^T enables a reverse rotation of the rigid body.

*Roughly putting, if you hold a bold object such as a metal ball and rotate your arm, that is a rotation of a rigid body, and your shoulder is the origin point. On the other hand, if you hold something soft like a marshmallow, it would be squashed in your hand, and that is not a not a rotation of a rigid body.

You can rotate \boldsymbol{x} with U like U^T\boldsymbol{x} = \left( \begin{array}{c} -\boldsymbol{u}_1^{T}- \\ -\boldsymbol{u}_2^{T}- \\ -\boldsymbol{u}_3^{T}- \end{array} \right)\boldsymbol{x}=\left( \begin{array}{c} \boldsymbol{u}_1^{T}\boldsymbol{x} \\ \boldsymbol{u}_2^{T}\boldsymbol{x} \\ \boldsymbol{u}_3^{T}\boldsymbol{x} \end{array} \right), and \boldsymbol{u}_i^{T}\boldsymbol{x} is the coordinate of \boldsymbol{x} projected on the axis \boldsymbol{u}_i.

Let’s see this more visually. Assume that the data point \boldsymbol{x}  is a purple dot and its position is expressed in the original coordinate system spanned by black arrows . By multiplying \boldsymbol{x} with U^T, the purple point \boldsymbol{x} is projected on the red axes respectively, and the product \left( \begin{array}{c} \boldsymbol{u}_1^{T}\boldsymbol{x} \\ \boldsymbol{u}_2^{T}\boldsymbol{x} \\ \boldsymbol{u}_3^{T}\boldsymbol{x} \end{array} \right) denotes the coordinate point of the purple point in the red coordinate system. \boldsymbol{x} is rotated this way, but for now I think it is better to think that the data are projected on new coordinate axes rather than the data themselves are rotating.

Now that we have seen what rotation by U means, you should have clearer image on what \left( \begin{array}{c} \xi \\ \eta \\ \zeta \end{array} \right)  = U^T \left( \begin{array}{c} x - \bar{x} \\ y - \bar{y} \\ z - \bar{z} \end{array} \right) means. \left( \begin{array}{c} \xi \\ \eta \\ \zeta \end{array} \right) denotes the coordinates of data projected on new axes \boldsymbol{u}_1, \boldsymbol{u}_2, \boldsymbol{u}_3, which are unit eigenvectors of V_{xyz}. In the coordinate system spanned by the eigenvectors, the data distribute like below.

By multiplying U from both sides of the equation above, we get \left( \begin{array}{c} x - \bar{x} \\ y - \bar{y} \\ z - \bar{z} \end{array} \right) =U \left( \begin{array}{c} \xi \\ \eta \\ \zeta \end{array} \right), which means you can express deviations of the original data as linear combinations of the three factors \xi, \eta, and \zeta. We expect that those three factors contain keys for understanding the original data more efficiently. If you concretely write down all the equations for the factors: \xi = 0.540 (x - \bar{x}) + 0.602 (y - \bar{y}) + 0.588 (z - \bar{z}), \eta = 0.736(x - \bar{x}) - 0.677 (y - \bar{y}) + 0.0174 (z - \bar{z}), and \zeta = - 0.408 (x - \bar{x}) - 0.423 (y - \bar{y}) + 0.809(z - \bar{z}). If you examine the coefficients of the deviations (x - \bar{x}), (y - \bar{y}), and (z - \bar{z}), we can observe that \eta almost equally reflects the deviation of the scores of all the subjects, thus we can say \eta is a factor indicating one’s general academic level. When it comes to \eta Japanese and Math scores are important, so we can guess that this factor indicates whether the student is at more of “scientific side” or “liberal art side.” In the same way \zeta relatively makes much of one’s English score,  so it should show one’s “internationality.” However the covariance of the data \xi, \eta, \zeta is V_{\xi \eta \zeta} = \begin{pmatrix} 148.34 & 0 & 0 \\ 0 & 30.62 & 0 \\ 0 & 0 & 3.60 \end{pmatrix}. You can see \zeta does not vary from students to students, which means it is relatively not important to describe the tendency of data. Therefore for dimension reduction you can cut off the factor \zeta.

*Assume that you can apply PCA on D-dimensional data and that you get \boldsymbol{x}', where \boldsymbol{x}' = U^T\boldsymbol{x} - \bar{\boldsymbol{x}}. The variance of data projected on new D-dimensional coordinate system is V'=\frac{1}{N}\sum{(\boldsymbol{x}')^T\boldsymbol{x}'} =\frac{1}{N}\sum{(U^T\boldsymbol{x})^T(U^T\boldsymbol{x})} =\frac{1}{N}\sum{U^T\boldsymbol{x}\boldsymbol{x}^TU} =U^T(\frac{1}{N}\sum{\boldsymbol{x}\boldsymbol{x}^T})U =U^TVU =diag(\lambda_1, \dots, \lambda_D). This means that in the new coordinate system after PCA, covariances between any pair of variants are all zero.

*As I mentioned U is a rotation of a rigid body, and U^T is the reverse rotation, hence U^TU = UU^T = I.

Hence you can approximate the original 3 dimensional data on the coordinate system (\boldsymbol{e}_1, \boldsymbol{e}_2, \boldsymbol{e}_3) from the reduced two dimensional coordinate system (\boldsymbol{u}_1, \boldsymbol{u}_2) with the following equation: \left( \begin{array}{c} x - \bar{x} \\ y - \bar{y} \\ z - \bar{z} \end{array} \right) \approx U_{reduced} \left( \begin{array}{c} \xi \\ \eta  \end{array} \right)  = (\boldsymbol{u}_1 \quad \boldsymbol{u}_2) \left( \begin{array}{c} \xi \\ \eta  \end{array} \right). Then it mathematically clearer that we can express the data with two factors: “how smart the student is” and “whether he is at scientific side or liberal art side.”

We can observe that eigenvalue \lambda_i is a statistic which indicates how much the corresponding \boldsymbol{u}_i can express the data, \frac{\lambda_i}{\sum_{j=1}^{D}{\lambda_j}} is called the contribution ratio of eigenvector \boldsymbol{u}_i. In the example above, the contribution ratios of \boldsymbol{u}_1, \boldsymbol{u}_2, and \boldsymbol{u}_3 are respectively \frac{\lambda_1}{\lambda_1 + \lambda_2 + \lambda_3}=0.813, \frac{\lambda_2}{\lambda_1 + \lambda_2 + \lambda_3}=0.168, \frac{\lambda_3}{\lambda_1 + \lambda_2 + \lambda_3}=0.0197. You can decide how many degrees of dimensions you reduce based on this information.

Appendix: Playing with my toy PCA on MNIST dataset

Applying “so called” PCA on MNIST dataset is a super typical topic that many other tutorial on PCA also introduce, but I still recommend you to actually implement, or at least trace PCA implementation with MNIST dataset without using libraries like scikit-learn. While reading this article I recommend you to actually run the first and the second code below. I think you can just copy and paste them on your tool to run Python, installing necessary libraries. I wrote them on Jupyter Notebook.

In my implementation, in the simple configuration part you can set the USE_ALL_NUMBERS as True or False boolean. If you set it as True, you apply PCA on all the data of numbers from 0 to 9. If you set it as True, you can specify which digit to apply PCA on. In this article, I show the results results of PCA on the data of digit ‘3.’ The first three images of ‘3’ are as below.

You have to keep it in mind that the data are all shown as 28 by 28 pixel grayscale images, but in the process of PCA, they are all processed as 28 * 28 = 784 dimensional vectors. After applying PCA on the 784 dimensional vectors of images of ‘3,’ the first 25 eigenvectors are as below. You can see that at the beginning the eigenvectors partly retain the shapes of ‘3,’ but they are distorted as the eigenvalues get smaller. We can guess that the latter eigenvalues are not that helpful in reconstructing the shape of ‘3.’

Just as we saw in the last section, you you can cut off axes of eigenvectors with small eigenvalues and reduce the dimension of MNIST data. The figure below shows how contribution ratio of MNIST data grows. You can see that around 200 dimension degree, the contribution ratio reaches around 0.95. Then we can guess that even if we reduce the dimension of MNIST from 784 to 200 we can retain the most of the structure of original data.

Some results of reconstruction of data from 200 dimensional space are as below. You can set how many images to display by adjusting NUMBER_OF_RESULTS in the code. And if you set LATENT_DIMENSION as 784, you can completely reconstruct the data.

* I make study materials on machine learning, sponsored by DATANOMIQ. I do my best to make my content as straightforward but as precise as possible. I include all of my reference sources. If you notice any mistakes in my materials, including grammatical errors, please let me know (email: yasuto.tamura@datanomiq.de). And if you have any advice for making my materials more understandable to learners, I would appreciate hearing it.

*I attatched the codes I used to make the figures in this article. You can just copy, paste, and run, sometimes installing necessary libraries.

 

 

Rethinking linear algebra: visualizing linear transformations and eigenvectors

In terms of calculation processes of Principal Component Analysis (PCA) or Linear Discriminant Analysis (LDA), which are the dimension reduction techniques I am going to explain in the following articles, diagonalization is what they are all about. Throughout this article, I would like you to have richer insight into diagonalization in order to prepare for understanding those basic dimension reduction techniques.

When our professor started a lecture on the last chapter of our textbook on linear algebra, he said “It is no exaggeration to say that everything we have studied is for this ‘diagonalization.'” Until then we had to write tons of numerical matrices and vectors all over our notebooks, calculating those products, adding their rows or columns to other rows or columns, sometimes transposing the matrices, calculating their determinants.

It was like the scene in “The Karate Kid,” where the protagonist finally understood the profound meaning behind the prolonged and boring “wax on, wax off” training given by Miyagi (or “jacket on, jacket off” training given by Jackie Chan). We had finally understood why we had been doing those seemingly endless calculations.

Source: http://thinkbedoleadership.com/secret-success-wax-wax-off/

But usually you can do those calculations easily with functions in the Numpy library. Unlike Japanese college freshmen, I bet you are too busy to reopen textbooks on linear algebra to refresh your mathematics. Thus I am going to provide less mathematical and more intuitive explanation of diagonalization in this article.

*This is the second article of the article series ” Illustrative introductions on dimension reduction .”

1, The mainstream ways of explaining diagonalization.

*The statements below are very rough for mathematical topics, but I am going to give priority to offering more visual understanding on linear algebra in this article. For further understanding, please refer to textbooks on linear algebra. If you would like to have minimum understandings on linear algebra needed for machine learning, I recommend the Appendix C of Pattern Recognition and Machine Learning by C. M. Bishop.

In most textbooks on linear algebra, the explanations on dioagonalization is like this (if you are not sure what diagonalization is or if you are allergic to mathematics, you do not have to read this seriously):

Let V (dimV = D)be a vector space and let  T_A : V \rightarrow V be a mapping of V into itself,  defined as T_A(v) = A \cdot \boldsymbol{v}, where A is a D\times D matrix and \boldsymbol{v} is D dimensional vector. An element \boldsymbol{v} \in V is called an eigen vector if there exists a number \lambda such that A \cdot \boldsymbol{v}= \lambda \cdot \boldsymbol{v} and \boldsymbol{v} \neq \boldsymbol{0}. In this case \lambda is uniquely determined and is called an eigen value of A belonging to the eigen vector \boldsymbol{v}.

Any matrix A has D eigen values \lambda_{i}, belonging to \boldsymbol{v}_{i} (i=1, 2, …., D). If \boldsymbol{v}_{i} is basis of the vector space V, then A is diagonalizable.

When A is diagonalizable, with D \times D matrices P = (\boldsymbol{v}_{1}, \dots, \boldsymbol{v}_{D}) , whose column vectors are eigen vectors \boldsymbol{v}_{i} (i=1, 2, …., D), the following equation holds: P^{-1}AP = \Lambda, where \Lambda = diag(\lambda_{1}, \dots, \lambda_{D})= \begin{pmatrix} \lambda_{1} & 0& \ldots &0\\ 0 & \lambda_{2} & \ldots & 0 \\ \vdots & \vdots & \ddots & \vdots \\ 0 & 0 & \ldots & \lambda_{D} \end{pmatrix}.

And when A is diagonalizable, you can diagonalize A as below.

Most textbooks keep explaining these type of stuff, but I have to say they lack efforts to make it understandable to readers with low mathematical literacy like me. Especially if you have to apply the idea to data science field, I believe you need more visual understanding of diagonalization. Therefore instead of just explaining the definitions and theorems, I would like to take a different approach. But in order to understand them in more intuitive ways, we first have to rethink waht linear transformation T_A means in more visible ways.

2, Linear transformations

Even though I did my best to make this article understandable to people with little prerequisite knowledge, you at least have to understand linear transformation of numerical vectors and with matrices. Linear transformation is nothing difficult, and in this article I am going to use only 2 or 3 dimensional numerical vectors or square matrices. You can calculate linear transformation of \boldsymbol{v} by A as equations in the figure. In other words, \boldsymbol{u} is a vector transformed by A.

*I am not going to use the term “linear transformation” in a precise way in the context of linear algebra. In this article or in the context of data science or machine learning, “linear transformation” for the most part means products of matrices or vectors. 

*Forward/back propagation of deep learning is mainly composed of this linear transformation. You keep linearly transforming input vectors, frequently transforming them with activation functions, which are for the most part not linear transformation.

As you can see in the equations above, linear transformation with A transforms a vector to another vector. Assume that you have an original vector \boldsymbol{v} in grey and that the vector \boldsymbol{u} in pink is the transformed \boldsymbol{v} by A is. If you subtract \boldsymbol{v} from \boldsymbol{u}, you can get a displacement vector, which I displayed in purple. A displacement vector means the transition from a vector to another vector.

Let’s calculate the displacement vector with more vectors \boldsymbol{v}. Assume that A =\begin{pmatrix} 3 & 1 \\ 1 & 2 \end{pmatrix}, and I prepared several grid vectors \boldsymbol{v} in grey as you can see in the figure below. If you transform those grey grid points with A, they are mapped into the vectors \boldsymbol{u} in pink. With those vectors in grey or pink, you can calculate the their displacement vectors \boldsymbol{u} - \boldsymbol{v}, which are in purple.

The displacement vectors in the figure above have some tendencies. In order to see that more clearly, let’s calculate displacement vectors with several matrices A and more grid points. Assume that you have three 2 \times 2 square matrices A_1 =\begin{pmatrix} 3 & 1 \\ 1 & 2 \end{pmatrix}, A_2 =\begin{pmatrix} 3 & 1 \\ -1 & 1 \end{pmatrix}, A_3 =\begin{pmatrix} 1 & -1 \\ 1 & 1 \end{pmatrix}, and I plotted displace vectors made by the matrices respectively in the figure below.

I think you noticed some characteristics of the displacement vectors made by those linear transformations: the vectors are swirling and many of them seem to be oriented in certain directions. To be exact, some displacement vectors extend in the same directions as some of original vectors in grey. That means  linear transformation by A did not change the direction of the original vector \boldsymbol{v}, and the unchanged vectors are called eigen vectors. Real eigen vectors of each A are displayed as arrows in yellow in the figure above. But when it comes to A_3, the matrix does not have any real eigan values.

In linear algebra, depending on the type matrices A, you have to consider various cases such as whether the matrices have real or imaginary eigen values, whether the matrices are diagonalizable, whether the eigen vectors are orthogonal, or whether they are unit vectors. But those topics are out of the scope of this article series, so please refer to textbooks on linear algebra if you are interested.

Luckily, however, in terms of PCA or LDA, you only have to consider a type of matrices named positive semidefinite matrices, which A_1 is classified to, and I am going to explain positive semidefinite matrices in the fourth section.

3, Eigen vectors as coordinate system

Source: Ian Stewart, “Professor Stewart’s Cabinet of Mathematical Curiosities,” (2008), Basic Books

Let me take Fibonacci numbers as an example to briefly see why diagonalization is useful. Fibonacci is sequence is quite simple and it is often explained using an example of pairs of rabbits increasing generation by generation. Let a_n (n=0, 1, 2, …) be the number of pairs of grown up rabbits in the n^{th} generation. One pair of grown up rabbits produce one pair of young rabbit The concrete values of a_n are a_0 = 0, a_1 = 1, a_2=1, a_3=2, a_4=3, a_5=5, a_6=8, a_7=13, \dots. Assume that A =\begin{pmatrix} 1 & 1 \\ 1 & 0 \end{pmatrix} and that \begin{pmatrix} a_1 \\ a_0  \end{pmatrix} =\begin{pmatrix} 1 \\ 0  \end{pmatrix}, then you can calculate the number of the pairs of grown up rabbits in the next generation with the following recurrence relation. \begin{pmatrix} a_{n+1} \\ a_{n}  \end{pmatrix}=\begin{pmatrix} 1 & 1 \\ 1 & 0 \end{pmatrix} \cdot \begin{pmatrix} a_{n+1} \\ a_{n}  \end{pmatrix}.Let \boldsymbol{a}_n be \begin{pmatrix} a_{n+1} \\ a_{n}  \end{pmatrix}, then the recurrence relation can be written as \boldsymbol{a}_{n+1} = A \boldsymbol{a}_n, and the transition of \boldsymbol{a}_n are like purple arrows in the figure below. It seems that the changes of the purple arrows are irregular if you look at the plots in normal coordinate.

Assume that \lambda _1, \lambda_2 (\lambda _1< \lambda_2) are eigen values of A, and \boldsymbol{v}_1, \boldsymbol{v}_2 are eigen vectors belonging to them respectively. Also let \alpha, \beta scalars such that \begin{pmatrix} a_{1} \\ a_{0}  \end{pmatrix} = \begin{pmatrix} 1 \\ 0  \end{pmatrix} = \alpha \boldsymbol{v}_1 + \beta \boldsymbol{v}_2. According to the definition of eigen values and eigen vectors belonging to them, the following two equations hold: A\boldsymbol{v}_1 = \lambda_1 \boldsymbol{v}_1, A\boldsymbol{v}_2 = \lambda_2 \boldsymbol{v}_2. If you calculate \boldsymbol{a}_1 is, using eigen vectors of A, \boldsymbol{a}_1  = A\boldsymbol{a}_0 = A (\alpha \boldsymbol{v}_1 + \beta \boldsymbol{v}_2) = \alpha\lambda _1 \boldsymbol{v}_1 + \beta \lambda_2 \boldsymbol{v}_2. In the same way, \boldsymbol{a}_2 = A\boldsymbol{a}_1 = A (\alpha\lambda _1 \boldsymbol{v}_1 + \beta \lambda_2 \boldsymbol{v}_2) = \alpha\lambda _{1}^{2} \boldsymbol{v}_1 + \beta \lambda_{2}^{2} \boldsymbol{v}_2, and \boldsymbol{a}_3 = A\boldsymbol{a}_2 = A (\alpha\lambda _{1}^{2} \boldsymbol{v}_1 + \beta \lambda_{2}^{2} \boldsymbol{v}_2) = \alpha\lambda _{1}^{3} \boldsymbol{v}_1 + \beta \lambda_{2}^{3} \boldsymbol{v}_2. These equations show that in coordinate system made by eigen vectors of A, linear transformation by A is easily done by just multiplying eigen values with each eigen vector. Compared to the graph of Fibonacci numbers above, in the figure below you can see that in coordinate system made by eigen vectors the plots changes more systematically generation by generation.

 

In coordinate system made by eigen vectors of square matrices, the linear transformations by the matrices can be much more straightforward, and this is one powerful strength of eigen vectors.

*I do not major in mathematics, so I am not 100% sure, but vectors in linear algebra have more abstract meanings. Various things in mathematics can be vectors, even though in machine learning or data science we  mainly use numerical vectors with more concrete elements. We can also say that matrices are a kind of maps. That is just like, at least in my impression, even though a real town is composed of various components such as houses, smooth or bumpy roads, you can simplify its structure with simple orthogonal lines, like the map of Manhattan. But if you know what the town actually looks like, you do not have to follow the zigzag path on the map.

4, Eigen vectors of positive semidefinite matrices

In the second section of this article I told you that, even though you have to consider various elements when you discuss general diagonalization, in terms of PCA and LDA we mainly use only a type of matrices named positive semidefinite matrices. Let A be a D \times D square matrix. If \boldsymbol{x}^T A \boldsymbol{x} \geq 0 for all values of the vector \boldsymbol{x}, the A is said to be a positive semidefinite matrix. And also it is known that A being a semidefinite matrix is equivalent to \lambda _{i} \geq 0 for all the eigen values \lambda_i (i=1, \dots , D).

*I think most people first learn a type of matrices called positive definite matrices. Let A be aD \times D square matrix. If \boldsymbol{x}^T A \boldsymbol{x} > 0 for all values of the vector \boldsymbol{x}, the A is said to be a positive definite matrix. You have to keep it in mind that even if all the elements of A are positive, A is not necessarly positive definite/semidefinite.

Just as we did in the second section of this article, let’s visualize displacement vectors made by linear transformation with a 3 \times 3 square positive semidefinite matrix A.

*In fact A_1 =\begin{pmatrix} 3 & 1 \\ 1 & 2 \end{pmatrix}, whose linear transformation I visualized the second section, is also positive semidefinite.

Let’s visualize linear transformations by a positive definite matrix A = \frac{1}{50} \begin{pmatrix} 60.45 &  33.63 & 46.29 \\33.63 & 68.49 & 50.93 \\ 46.29 & 50.93 & 53.61 \end{pmatrix}. I visualized the displacement vectors made by the A just as the same way as in the second section of this article. The result is as below, and you can see that, as well as the displacement vectors made by A_1, the three dimensional displacement vectors below are swirling and extending in three directions, in the directions of the three orthogonal eigen vectors \boldsymbol{v}_1, \boldsymbol{v}_2, and \boldsymbol{v}_3.

*It might seem like a weird choice of a matrix, but you are going to see why I chose it in the next article.

You might have already noticed A_1 =\begin{pmatrix} 3 & 1 \\ 1 & 2 \end{pmatrix} and A = \frac{1}{50} \begin{pmatrix} 60.45 &  33.63 & 46.29 \\33.63 & 68.49 & 50.93 \\ 46.29 & 50.93 & 53.61 \end{pmatrix} are both symmetric matrices and that their elements are all real values, and that their diagonal elements are all positive values. Super importantly, when all the elements of a D \times D symmetric matrix A are real values and its eigen values are \lambda_{i} (i=1, \dots , D), there exist orthonormal matrices U such that U^{-1}AU = \Lambda, where \Lambda = diag(\lambda_{1}, \dots , \lambda_{D}).

*The title of this section might be misleading, but please keep it in mind that positive definite/semidefinite matrices are not necessarily real symmetric matrices. And real symmetric vectors are not necessarily positive definite/semidefinite matrices.

5, Orthonormal matrices and rotation of vectors

In this section I am gong to explain orthonormal matrices, as known as rotation matrices. If a D\times D matrix U is an orthonormal matrix, column vectors of U are orthonormal, which means U = (\boldsymbol{u}_1 \dots \boldsymbol{u}_D), where \begin{cases} \boldsymbol{u}_{i}^{T}\boldsymbol{u}_{j} = 1 \quad (i = j) \\ \boldsymbol{u}_{i}^{T}\boldsymbol{u}_{j} = 0 \quad (i\neq j) \end{cases}. In other words column vectors \boldsymbol{u}_{i} form an orthonormal coordinate system.

Orthonormal matrices U have several important properties, and one of the most important properties is U^{-1} = U^{T}. Combining this fact with what I have told you so far, you we can reach one conclusion: you can orthogonalize a real symmetric matrix A as U^{T}AU = \Lambda. This is known as spectral decomposition or singular value decomposition.

Another important property of U is that U^{T} is also orthonormal. In other words, assume U is orthonormal and that U = (\boldsymbol{u}_1 \dots \boldsymbol{u}_D) = \begin{pmatrix} -\boldsymbol{v_1}^{T}- \\ \vdots \\ -\boldsymbol{v_D}^{T}- \end{pmatrix}, (\boldsymbol{v}_1 \dots \boldsymbol{v}_D) also forms a orthonormal coordinate system.

…It seems things are getting too mathematical and abstract (for me), thus for now I am going to wrap up what I have explained in this article .

We have seen

  • Numerical matrices linearly transform vectors.
  • Certain linear transformations do not change the direction of vectors in certain directions, which are called eigen vectors.
  • Making use of eigen vectors, you can form new coordinate system which can describe the linear transformations in a more straightforward way.
  • You can diagonalize a real symmetric matrix A with an orthonormal matrix U.

Of our current interest is what kind of linear transformation the real symmetric positive definite matrix enables. I am going to explain why the purple vectors in the figure above is swirling in the upcoming articles. Before that, however, we are going to  see one application of what we have seen in this article, on dimension reduction. To be concrete the next article is going to be about principal component analysis (PCA), which is very important in many fields.

*In short, the orthonormal matrix U, which I mentioned above enables rotation of matrix, and the diagonal matrix diag(\lambda_1, \dots, \lambda_D) expands or contracts vectors along each axis. I am going to explain that more precisely in the upcoming articles.

* I make study materials on machine learning, sponsored by DATANOMIQ. I do my best to make my content as straightforward but as precise as possible. I include all of my reference sources. If you notice any mistakes in my materials, including grammatical errors, please let me know (email: yasuto.tamura@datanomiq.de). And if you have any advice for making my materials more understandable to learners, I would appreciate hearing it.

*I attatched the codes I used to make the figures in this article. You can just copy, paste, and run, sometimes installing necessary libraries.

 

Spiky cubes, Pac-Man walking, empty M&M’s chocolate: curse of dimensionality

This is the first article of the article series Illustrative introductions on dimension reduction.

“Curse of dimensionality” means the difficulties of machine learning which arise when the dimension of data is higher. In short if the data have too many features like “weight,” “height,” “width,” “strength,” “temperature”…., that can undermine the performances of machine learning. The fact might be contrary to your image which you get from the terms “big” data or “deep” learning. You might assume that the more hints you have, the better the performances of machine learning are. There are some reasons for curse of dimensionality, and in this article I am going to introduce two major reasons below.

  1. High dimensional data usually have rich expressiveness, but usually training data are too poor for that.
  2. The behaviors of data points in high dimensional space are totally different from our common sense.

Through these topics, you will see that you always have to think about which features to use considering the number of data points.

*From now on I am going to talk about only Euclidean distance. If you are not sure what Euclidean distance means, please just keep it in mind that it is the type of distance most people wold have learnt in normal compulsory education.

*This is the first article of the article series ” Illustrative introductions on dimension reduction .”

1. Number of samples and degree of dimension

The most straightforward demerit of adding many features, or increasing dimensions of data, is the growth of computational costs. More importantly, however, you always have to think about the degree of dimensions in relation of the number of data points you have. Let me take a simple example in a book “Pattern Recognition and Machine Learning” by C. M. Bishop (PRML). This is an example of measurements of a pipeline. The figure below shows a comparison plot of 3 classes (red, green and blue), with parameter x_7 plotted against parameter x_6 out of 12 parameters.

* The meaning of data is not important in this article. If you are interested please refer to the appendix in PRML.

Assume that we are interested in classifying the cross in black into one of the three classes. One of the most naive ideas of this classification is dividing the graph into grids and labeling each grid depending on the number of samples in the classes (which are colored at the right side of the figure). And you can classify the test sample, the cross in black, into the class of the grid where the test sample is in. Thereby the cross is classified to the class in red.

Source: C.M. Bishop, “Pattern Recognition and Machine Learning,” (2006), Springer, pp. 34-35

As I mentioned in the figure above, we used only two features out of 12 features in total. When the total number of data points is fixed and you add remaining ten axes/features one after another, what would happen? Let’s see what “adding axes/features” means. If you are talking about 1, 2, or 3 dimensional grids, you can visualize them. And as you can see from the figure below, if you make each 10^1, 10^2, 100^3 grids respectively in 1, 2, 3 dimensional spaces, the number of the small regions in the grids are respectively 10, 100, 1000. Even though you cannot visualize it anymore, you can make grids for more than 3 dimensional data. If you continue increasing the degree of dimension, the number of grids increases exponentially, and that can soon surpass the number of training data points. That means there would be a lot of empty spaces in such high dimensional grids. And the classifying method above: coloring each grid and classifying unknown samples depending on the colors of the grids, does not work out anymore because there would be a lot of empty grids.

* If you are still puzzled by the idea of “more than 3 dimensional grids,” you should not think too much about that now. It is enough if you can get some understandings on high dimensional data after reading the whole article of this.

Source: Goodfellow and Yoshua Bengio and Aaron Courville, Deep Learning, (2016), MIT Press, p. 153

I said the method above is the most naive way, but other classical classification methods , for example k-nearest neighbors algorithm, are more or less base on a similar idea. Many of classical machine learning algorithms are based on the idea of smoothness prior, or local constancy prior. In short in classical ways, you  do not expect data to change so much in a small region, so you can expect unknown samples to be similar to data in vicinity. But that soon turns out to be problematic when the dimension of data is bigger because training data would be sparse because the area of multidimensional space grows exponentially as I mentioned above. And sometimes you would not be able to find training data around test data. Plus, in high dimensional data, you cannot treat distance in the same as you do in lower dimensional space. The ideas of “close,” “nearby,” or “vicinity” get more obscure in high dimensional data. That point is related to the next topic: the intuition have cultivated in normal life is not applicable to higher dimensional data.

2. Bizarre characteristics of high dimensional data

We form our sense of recognition in 3-dimensional ways in our normal life. Even though we can visualize only 1, 2, or 3 dimensional data, we can actually generalize the ideas in 1, 2, or 3 dimensional ideas to higher dimensions. For example 4 dimensional cubes, 100 dimensional spheres, or orthogonality in 255 dimensional space. Again, you cannot exactly visualize those ideas, and for many people, such high dimensional phenomenon are just imaginary matters on blackboards. Those high dimensional ideas are designed to retain some conditions just as well as 1, 2, or 3 dimensional space. Let’s take an example of spheres in several dimensional spaces. General spheres in any D-dimensional space can be defined as a set of any \boldsymbol{x}, such that |\boldsymbol{x} - \boldsymbol{c}| = r, where \boldsymbol{c} is the center point and r is length of radius. When \boldsymbol{x} is 2-dimensional, the spheres are called “circles.” When \boldsymbol{x} is 3-dimensional, the spheres are called “spheres” in our normal life, unless it is used in a conversation in a college cafeteria, by some students in mathematics department. And when \boldsymbol{x} is D-dimensional, they are called D-ball, and again, this is just a imaginary phenomenon on blackboard.

* Vectors and points are almost the same because all the vectors are denoted as “arrows” from the an origin point to sample data points.  The only difference is that when you use vectors, you have to consider their directions.

* “D-ball” is usually called “n-ball,” and in such context it is a sphere in a n-dimensional space. But please let me use the term “D-ball” in this article.

Not only spheres, but only many other ideas have been generalized to D-dimensional space, and many of them are indispensable also for data science. But there is one severe problem: the behaviors of data in high dimensional field is quite different from those in two or three dimensional space. To be concrete, in high dimensional field, cubes are spiky, you have to move like Pac-Man, and M & M’s Chocolate looks empty inside but tastes normal.

2.1: spiky cubes
Let’s take a look at an elementary-school-level example of geometry first. Assume that you have several unit squares or unit cubes like below. In each of them a circle or sphere with diameter 1 is inscribed. The length of a diagonal line in each square is \sqrt{2}, and that in each cube is \sqrt{3}.

If you stack the squares or cubes as below, what are the length of diameters of the blue circle or sphere, circumscribing all the 4 orange circles or the 8 orange spheres?

The answers are, the diameter of the blue circle is \sqrt{2} - 1, and the diameter of the blue sphere is \sqrt{3} - 1.

Next let’s think about the same situation in higher dimensional space. Assume that there are some unit D-dimensional hypercubes stacked, in each of which a D-ball with diameter 1 is inscribed, touching all the surfaces inside. Then what is the length of the diameter of  a D-ball circumscribing all the unit D-ball in the hypercubes ? Given the results above, it ca be predicted that its diameter is \sqrt{D}  -1. If that is true, there is one strange point: \sqrt{D} - 1 can soon surpass 2: that means in the chart above the blue sphere will stick out of the stacked cubes. That sounds like a paradox, but with one hypothesis, the phenomenon makes sense: cubes become more spiky as the degree of dimension grows. This hypothesis is a natural deduction because diagonal lines of hyper cubes get longer, and the the center of each surface of hypercubes still touches the unit D-ball with diameter 1, inscribing inscribing inside each unit hypercube.

If you stack 4 hypercubes, the blue sphere circumscribing them will not stick out of the stacked hypercubes anymore like the figure below.

*Of course you cannot visualize what is going on in D-dimensional space, so the figure below is just a pseudo simulation of D-dimensional space in our 3-dimensional sense. I guess you have to stack more than four hyper cubes in higher dimensional data, but you cannot easily imagine what will go on in such space anymore.

 

*You can confirm the fact that hypercube gets more spiky as the degree of dimension growth, by comparing the volume of the hypercube and the volume of the D-ball inscribed inside the hypercube. Thereby you can prove that the volume of hypercube concentrates on the corners of the hypercube. Plus, as I mentioned the longest diagonal distance of hypercube gets longer as dimension degree increases. That is why hypercube is said to be spiky. For mathematical proof, please check the Exercise 1.19 of PRML.

2.2: Pac-Man walking

Next intriguing phenomenon in high dimensional field is that most of pairs of vectors in high dimensional space are orthogonal. In other words, if you select two random vectors in high dimensional space, the angle between them are mostly close to 90^\circ. Let’s see the general meaning of angle between two vectors in any dimensional spaces. Assume that the angle between two vectors \boldsymbol{u}, and \boldsymbol{v} is \theta, then cos\theta is calculated as cos\theta = \frac{<\boldsymbol{u}, \boldsymbol{v}>}{|\boldsymbol{u}||\boldsymbol{v}|}. In 1, 2, or 3 dimensional space, you can actually see the angle, but again you can define higher dimensional angle, which you cannot visualize anymore. And angles are sometimes used as similarity of two vectors.

* <\boldsymbol{u}, \boldsymbol{v}> is the inner product of \boldsymbol{u}, and \boldsymbol{v}.

Assume that you generate a pair of two points inside a D-dimensional unit sphere and make two vectors \boldsymbol{u}, and \boldsymbol{v} by connecting the origin point and those two points respectively. When D is 2, I mean spheres are circles in this case, any \theta are equally generated as in the chart below. The fact might be the same as your intuition.   How about in 3-dimensional space? In fact the distribution of \theta is not uniform. \theta = 90^\circ is the most likely to be generated. As I explain in the figure below, if you compare the area of cross section of a hemisphere and the area of a cone whose vertex is the center point of the sphere, you can see why.

I generated 10000 random pairs of points in side a D-dimensional unit sphere, and calculated the angle between them. In other words I just randomly generated two D-dimensional vectors \boldsymbol{u} and \boldsymbol{v}, whose elements are randomly generated values between -1 and 1, and calculated the angle between them, repeating this process 10000 times. The chart below are the histograms of angle between pairs of generated vectors in respectively 2, 3, 50, and 100 dimensional space.

As I explained above, in 2-dimensional space, the distribution of \theta is almost uniform. However the distribution concentrates a little around 90^\circ in 3-dimensional space. You can see that the bigger the degree of dimension is, the more the angles of generated vectors concentrate around 90^\circ. That means most pairs of vectors in high dimensional space are close to orthogonal. Movements are also sequence of vectors, so when most pairs of movement vectors are orthogonal, that means you can only move like Pac-Man in such space.

Source: https://edition.cnn.com/style/article/pac-man-40-anniversary-history/index.html

* Of course I am talking about arcade Mac-Man game. Not Pac-Man in Super Smash Bros.  Retro RPG video games might have more similar playability, but in high dimensional space it is also difficult to turn back. At any rate, I think you have understood it is even difficult to move smoothly in high dimensional space, just like the first notorious Resident Evil on the first PS console also had terrible playability .

2.3: empty M & M’s chocolate

Let’s think about the proportion of the volume of the outermost \epsilon surface of general spheres with radius r. First, in 2 two dimensional space, spheres are circles. The area of the brown part of the circle below is \pi r^2. In order calculate the are of \epsilon \cdot r thick surface of the circle, you have only to subtract the area of \pi \{ (1 - \epsilon)\cdot r\} ^2. When \epsilon = 0.01, the area of outer most surface is \pi r^2 - \pi (0.99\cdot r)^2, and its proportion to the area of the whole circle is \frac{\pi r^2 - \pi (0.99\cdot r)^2}{\pi r^2} = 0.0199.

In case of 3-dimensional space, the value of a sphere with radius r is \frac{4}{3} \pi r^2, so the proportion of the \epsilon surface is calculated in the same way: \frac{\frac{4}{3} \pi r^3 -\frac{4}{3} \pi (0.99\cdot r)^2}{\frac{4}{3}\pi r^2} = 0.0297. Compared to the case in 2 dimensional space, the proportion is a little bigger.

How about in D-dimensional space? We have seen that even in  D-dimensional space the surface of a sphere, I mean D-ball, can be defined as a set of any points whose distance from the center point is all r. And it is known that the volume of D-ball is defined as below.

\Gamma () is called gamma function, but in this article it is not so important. The most important point now is, if you discuss any D-ball, their volume only depends on their radius r. That meas the proportion of outer \epsilon surface of D-ball is calculated as \frac{\pi r^2 - \pi \{ (1 - \epsilon)\cdot r\} ^2}{\pi r^2}. When \epsilon is 0.01, the proportion of the 1% surface of D-ball changes like in the chart below.

* And of course when D is 2,  \frac{\pi ^{(\frac{D}{2})}}{\Gamma (\frac{D}{2} + 1)} = \pi, and when D is 3 ,  \frac{\pi ^{(\frac{D}{2})}}{\Gamma (\frac{D}{2} + 1)} = \frac{4}{3} \pi

You can see that when D is over 400, around 90% of volume is concentrated in the very thin 1% surface of D-ball. That is why, in high dimensional space, M & M’s chocolate look empty but tastes normal: all the chocolate are concentrated beneath the sugar coating.

More interestingly, even if you choose any points as a central point of a sphere with radius r, the other points are squashed to the surface of the sphere, even if all the data points are uniformly distributed. This situation is problematic for classical machine learning algorithms, which are often based on the Euclidean distances between pairs of two sample data points: if you go from the central point to another sample point, the possibility of finding the point within (1 - \epsilon)\cdot r radius of the center is almost zero. But if you reach the outermost \epsilon part of the surface of the sphere, most data points are there. However, for one of the data points in the surface, any other data points are distant in the same way.

Inside M & M’s chocolate is a mysterious world.

Source: https://hipwallpaper.com/mms-wallpapers/

You have seen that using high dimensional data can be problematic in many ways. Data science and machine learning are largely based on one idea: you can find a lower dimensional meaningful and easier structure in data. In the next articles I am going to introduce some famous dimension reduction algorithms. And hopefully I would like to give some deeper insights in to these algorithms, in straightforward ways.

* I could not explain the relationships of variance and bias of data. This is also a very important factor when you think about dimensionality of data. I hope I can write about this topic someday. You can also look it up if you are interested.

[References]

[1]C. M. Bishop, “Pattern Recognition and Machine Learning,” (2006), Springer, pp. 33-37

[2]Goodfellow and Yoshua Bengio and Aaron Courville, Deep Learning, (2016), MIT Press, p. 153

[3] Shiga Kouji, “30 Lesson to Topology,” (1988)

[4]”Volume of an n-ball,” Wikipedia
https://en.wikipedia.org/wiki/Volume_of_an_n-ball

* I make study materials on machine learning, sponsored by DATANOMIQ. I do my best to make my content as straightforward but as precise as possible. I include all of my reference sources. If you notice any mistakes in my materials, including grammatical errors, please let me know (email: yasuto.tamura@datanomiq.de). And if you have any advice for making my materials more understandable to learners, I would appreciate hearing it.

Illustrative introductions on dimension reduction

“What is your image on dimensions?”

….That might be a cheesy question to ask to reader of Data Science Blog, but most people, with no scientific background, would answer “One dimension is a line, and two dimension is a plain, and we live in three-dimensional world.” After that if you ask “How about the fourth dimension?” many people would answer “Time?”

You can find books or writings about dimensions in various field. And you can use the word “dimension” in normal conversations, in many contexts.

*In Japanese, if you say “He likes two dimension.” that means he prefers anime characters to real women, as is often the case with Japanese computer science students.

The meanings of “dimensions” depend on the context, but in data science dimension is usually the number of rows of your Excel data.

When you study data science or machine learning, usually you should start with understanding the algorithms with 2 or 3 dimensional data, and you can apply those ideas to any D dimensional data. But of course you cannot visualize D dimensional data anymore, and you always have to be careful of what happens if you expand degree of dimension.

Conversely it is also important to reduce dimension to understand abstract high dimensional stuff in 2 or 3 dimensional space, which are close to our everyday sense. That means dimension reduction is one powerful way of data visualization.

In this blog series I am going to explain meanings of dimension itself in machine learning context and algorithms for dimension reductions, such as PCA, LDA, and t-SNE, with 2 or 3 dimensional visible data. Along with that, I am going to delve into the meaning of calculations so that you can understand them in more like everyday-life sense.

This article series is going to be roughly divided into the contents below.

  1. Curse of Dimensionality
  2. Rethinking linear algebra: visualizing linear transformations and eigen vector
  3. The algorithm known as PCA and my taxonomy of linear dimension reductions
  4. Rethinking linear algebra part two: ellipsoids in data science
  5. Autoencoder as dimension reduction (to be published soon)
  6. t-SNE (to be published soon)

I hope you could see that reducing dimension is one of the fundamental approaches in data science or machine learning.

Process Mining mit Celonis – Artikelserie

Der erste Artikel dieser Artikelserie Process Mining Tools beschäftigt sich mit dem Anbieter Celonis. Das 2011 in Deutschland gegründete Unternehmen ist trotz wachsender Anzahl an Wettbewerbern zum Zeitpunkt der Veröffentlichung dieses Artikels der eindeutige Marktführer im Bereich Process Mining.

Celonis Process Mining – Teil 1 der Artikelserie

Celonis Process Mining ist 2011 als reine On-Premise-Lösung gestartet und seit 2018 auch als Cloud-Lösuung zu haben. Übersicht zu den vier verschiedenen Produktversionen der Celonis Process Mining Lösungen:

Celonis Snap Celonis Enterprise Celonis Academic Celonis Consulting
Lizenz:  Kostenfrei Kostenpflichtige Lösungspakete Kostenfrei Consulting Lizenz on Demand
Zielgruppe:  Für kleine Unternehmen und Einzelanwender Für mittel- und große Unternehmen Für akademische Einrichtungen und Studenten Für Berater
Datenquellen: ServiceNow, CSV/XLS -Datei Beliebig (On-Premise- und Cloud – Anbindungen) ServiceNow, CSV/XLS/XES –Datei oder Demosysteme Beliebig (On-Premise- und Cloud – Anbindungen)
Datenvolumen: Limitiert auf 500 MB Event-Log-Daten Unlimitierte Datenmengen (Größte Installation 50 TB) Unlimitierte Datenmengen Unlimitierte Datenmengen (Größte Installation 30 TB
Architektur: Cloud & On-Premise Cloud & On-Premise Cloud & On-Premise Cloud & On-Premise

Dieser Artikel bezieht sich im weiteren Verlauf auf die Celonis Enterprise Version, wenn nicht anders gekennzeichnet. Spezifische Unterschiede unter den einzelnen Produkten und weitere Informationen können auf der Website von Celonis entnommen werden.

Bedienbarkeit und Anpassungsfähigkeit der Analysen

In Sachen Bedienbarkeit punktet Celonis mit einem sehr übersichtlichen und einsteigerfreundlichem Userinterface. Jeder der mit BI-Tools wir z.B. „Power-BI“ oder „Tableau“ gearbeitet hat, wird sich wahrscheinlich schnell zurechtfinden.

Userinterface Celonis

Abbildung 1: Userinterface von Celonis. Über die Reiter kann direkt von der Analyse (Process Analytics) zu den ETL-Prozessen (Event Collection) gewechselt werden.

Das Erstellen von Analysen funktioniert intuitiv und schnell, auch weil die einzelnen Komponentenbausteine lediglich per drag & drop platziert und mit den gewünschten Dimensionen und KPI’s bestückt werden müssen.

Process Analytics im Process Explorer

Abbildung 2: Typische Analyse im Edit Modus. Neue Komponenten können aus dem Reiter (rechts im Bild) mittels drag & drop auf der Dashboard Bearbeitungsfläche platziert werden.

Darüber hinaus bietet Celonis mit seinem kostenlosen Programm „Celonis Acadamy“ einen umfangreichen und leicht verständlichen Pool an Trainingseinheiten für die verschiedenen User-Rollen: „Snap“, „Executive“, „Business User“, „Analyst“ und „Data Engineer“. Einsteiger finden sich nach der Absolvierung der Grundkurse etwa nach vier Stunden in dem Tool zurecht.

Conformance Analyse In Celonis

Abbildung 3: Conformance Analyse In Celonis. Es kann direkt analysiert werden, welche Art von Verstößen welche Auswirkungen haben und mit welcher Häufigkeit diese auftreten.

Die Definition von eigenen KPIs erfolgt mittels übersichtlichem Code Editor. Die verwendete proprietäre und patentierte Programmiersprache lautet PQL (Process Query Language) , dessen Syntax stark an SQL angelehnt ist und alle prozessrelevanten Berechnungen ermöglicht. Noch einsteigerfreundlicher ist der Visual Editor, in welchem KPIs alternativ mit zahlreicher visueller Unterstützung und über 130 mathematischen Operatoren erstellt werden können – ganz ohne Coding Erfahrung.
Mit Hilfe von über 30 Komponenten lassen sich alle üblichen Charts und Grafiken erstellen. Ich hatte das Gefühl, dass die Auswahl grundsätzlich ausreicht und dem Erkenntnisgewinn nicht im Weg steht. Dieses Gefühl rührt nicht zuletzt daher, dass die vorgefertigten Features, wie zum Beispiel „Conformance“ direkt und ohne Aufwand implementiert werden können und bemerkenswerte Erkenntnisse liefern. Kurzum: Ja es ist vieles vorgefertigt, aber hier wurde mit hohen Qualitätsansprüchen vorgefertigt!

Celonis Code Editor vs Visual Editor

Abbildung 4: Coder Editor (links) und Visual Editor (rechts). Während im Code Editor mit PQL geschrieben werden muss, können Einsteiger im Visual Editor visuelle Hilfestellungen nehmen, um KPIs zu definieren.

Diese Flexibilität erscheint groß und bedient mehrere Zielgruppen, beginnend bei den Einsteigern. Insbesondere da das Verständnis für den Code Editor und somit für PQL durch die Arbeit mit dem Visual Code Editor gefördert wird. Wer SQL-Kenntnisse mitbringt, wird sehr schnell ohne Probleme KPIs im Code Editor definieren können. Erfahrenen Data Engineers stünde es dennoch frei, die Entwicklungsarbeit auf die Datenbankebene zu verschieben.

Celonis Visual Editor

Abbildung 5: Mit Hilfe zahlreicher Möglichkeiten können Einsteiger im Visual Editor visuelle Hilfestellungen nehmen, um individuelle KPIs zu definieren.

Nachdem die ersten Analysen erstellt wurden, steht der Prozessanalyse nichts mehr im Wege. Während sich per Knopfdruck auf alle visualisierten Datenpunkte filtern lässt, unterstützt auch hier Celonis zusätzlich mit zahlreichen sogenannten ‘Auswahlansichten’, um die Entdeckung unerwünschter oder betrügerischer Prozesse so einfach wie das Googeln zu machen.

Predefined dashboard apps

Abbildung 6: Die anwenderfreundlichen Auswahlarten ermöglichen es dem Benutzer, einfach mit wenigen Klicks nach Unregelmäßigkeiten oder Mustern in Transaktionen zu suchen und diese eingehend zu analysieren.

Integrationsfähigkeit

Die Celonis Enterprise Version ist sowohl als Cloud- und On-Premise-Lösung verfügbar. Die Cloud-Lösung bietet die folgenden Vorteile: Zum einen zusätzliche Leistungen wieCloud Connectoren, einer sogenannten Action Engine die jeden einzelnen Mitarbeiter in einem Unternehmen mit datengetriebenen nächstbesten Handlungen unterstützt, intelligenter Process Automation, Machine Learning und AI, einen App Store sowie verschiedene Boards. Diese Erweiterungen zeigen deutlich den Anspruch des Münchner Process Mining Vendors auf, neben der reinen Prozessanalyse Unternehmen beim heben der identifizierten Potentiale tatkräftig zu unterstützen. Darüber hinaus kann die Cloud-Lösung punkten mit, einer schnellen Amortisierung, bedarfsgerechter Skalierbarkeit der Kapazitäten sowie einen noch stärkeren Fokus auf Security & Compliance. Darüber hinaus  erfolgen regelmäßig Updates.

Celonis Process Automation

Abbildung 7: Celonis Process Automation ermöglicht Unternehmen ihre Prozesse auf intelligente Art und Weise so zu automatisieren, dass die Zielerreichung der jeweiligen Fachabteilung im Fokus stehen. Auch hier trumpft Celonis mit über 30+ vorgefertigten Möglichkeiten von der Automatisierung von Kommunikation, über Backend Automatisierung in Quellsystemen bis hin zu Einbindung von RPA Bots und vielem mehr.

Der Schwenk von Celonis scheint in Richtung Cloud zu sein und es bleibt abzuwarten, wie die On-Premise-Lösung zukünftig aussehen wird und ob sie noch angeboten wird. Je nach Ausgangssituation gilt es hier abzuwägen, welche der beiden Lösungen die meisten Vorteile bietet. In jedem Fall wird Celonis als browserbasierte Webanwendung für den Endanwender zur Verfügung gestellt. Die folgende Abbildung zeigt eine beispielhafte Celonis on-Premise-Architektur, bei welcher der User über den Webbrowser Zugang erhält.

Celonis bringt eine ausreichende Anzahl an vordefinierten Datenschnittstellen mit, wodurch sowohl gängige on-Premise Datenbanken / ERP-Systeme als auch Cloud-Dienste, wie z. B. „ServiceNow“ oder „Salesforce“ verbunden werden können. Im „App Store“ können zusätzlich sogenannte „prebuild Process-Connectors“ kostenlos erworben werden. Diese erstellen die Verbindung und erzeugen das Datenmodell (Extract and Transform) für einen Standard Prozess automatisch, so dass mit der Analyse direkt begonnen werden kann. Über 500 vordefinierte Analysen für Standard Prozesse gibt es zusätzlich im App Store. Dadurch kann die Bearbeitungszeit für ein Process-Mining Projekt erheblich verkürzt werden, vorausgesetzt das benötigte Datenmodel weicht im Kern nicht zu sehr von dem vordefinierten Model ab. Sollten Schnittstellen mal nicht vorhanden sein, können Daten auch als CSV oder XLS Format importiert werden.

Celonis App Store

Abbildung 8: Der Celonis App Store beinhaltet über 100 Prozesskonnektoren, über 500 vorgefertigte Analysen und über 80 Action Engine Fähigkeiten die kostenlos mit der Cloud Lizenz zur Verfügung stehen

Auch wenn von einer 100%-Cloud gesprochen wird, muss für die Anbindung von unternehmensinternen on-premise Datenquellen (z. B. lokale Instanzen von SAP ERP, Oracle ERP, MS Dynamics ERP) ein sogenannter Extractor on-premise installiert werden.

Celonis Extractors

Abbildung 9: Celonis Extractor muss für die Anbindung von On-Premise Datenquellen ebenfalls On-Premise installiert werden. Dieser arbeitet wie ein Gateway zur Celonis Intelligent Business Cloud (IBC). Die IBC enthält zudem einen eigenen Extratctor für die Anbindung von Daten aus anderen Cloud-Systemen.

Celonis bietet in der Enterprise-Ausführung zudem ein umfassendes Benutzer-Berechtigungsmanagement, so dass beispielsweise für Analysen im Einkauf die Berechtigungen zwischen dem Einkaufsleiter, Einkäufern und Praktikanten im Einkauf unterschieden werden können. Auch dieser Punkt ist für viele Unternehmen eine Grundvoraussetzung für einen eventuellen unternehmensweiten Roll-Out.

Skalierbarkeit

In Punkto großen Datenmengen kann Celonis sich sehen lassen. Allein für „Uber“ verarbeitet die Cloud rund 50 Millionen Datensätze, wobei ein einzelner mehrere Terabyte (TB) groß sein kann. Der größte einzelne Datenblock, den Celonis analysiert, beträgt wohl etwas über 50 TB. Celonis bietet somit Process Mining, zeitgerecht im Bereich Big Data an und kann daher auch viele große renommierten Unternehmen zu seinen Kunden zählen, wie zum Beispiel Siemens, ABB oder BMW. Doch wie erweiterbar und flexibel sind die erstellten Datenmodelle? An diesem Punkt konnte ich keine Schwierigkeiten feststellen. Celonis bietet ein übersichtlich gestaltetes Userinterface, welches das Datenmodell mit seinen Tabellen und Beziehungen sauber darstellt. Modelliert wird mit SQL-Befehlen, wodurch eine zusätzliche Abfragesprache entfällt. Der von Celonis gewählte SQL-Dialekt ist Vertica. Dieser ist keineswegs begrenzt und bietet die ausreichende Tiefe, welche an dieser Stelle benötigt wird. Die Erweiterbarkeit sowie die Flexibilität der Datenmodelle wird somit ausschließlich von der Arbeit des Data Engineer bestimmt und in keiner Weise durch Celonis selbst eingeschränkt. Durch das Zurückgreifen auf die Abfragesprache SQL, kann bei der Modellierung auf eine sehr breite Community zurückgegriffen werden. Darüber hinaus können bestehende SQL-Skripte eingefügt und leicht angepasst werden. Und auch die Suche nach einem geeigneten Data Engineer gestaltet sich dadurch praktisch, da SQL eine der meistbeherrschten Abfragesprachen ist.

Zukunftsfähigkeit

Machine Learning umfasst Data Mining und Predictive Analytics und findet vermehrt den Einzug ins Process Mining. Auch ist es längst ein wesentlicher Bestandteil von Celonis. So basiert z. B. das Feature „Conformance“ auf Machine Learning Algorithmen, welche zu den identifizierten Prozessabweichungen den Einfluss auf das Geschäft berechnen. Aber auch Lösungen zu den Identifizierten Problemen werden von Verfahren des maschinellen Lernens dem Benutzer vorgeschlagen. Was zusätzlich in Sachen Machine Learning von Celonis noch bereitgestellt wird, ist die sogenannte Machine-Learning-Workbench, welche in die Intelligent Business Cloud integriert ist. Hier können eigene Anwendungen mit Machine Learning auf Basis der Event-Log Daten entwickelt und eingesetzt werden, um z. B. Vorhersagen zu Lieferzeiten treffen zu können.

Task Mining ist einer der nächsten Schritte im Bereich Process Mining, der den Detailgrad für Analysen von Prozessen bis hin zu einzelnen Aufgaben auf Mausklick-Ebene erhöht. Im Oktober 2019 hatte Celonis bereits angekündigt, dass die Intelligent Business Cloud um eben diese neue Technik der Datenerhebung und -analyse erweitert wird. Die beiden Methoden Prozess Analyse und Task Mining ergänzen sich ausgezeichnet. Stelle ich in der Prozess Analyse fest, dass sich eine bestimmte Aktivität besonders negativ auf meine gewünschte Performance auswirkt (z. B. Zeit), können mit Task Mining diese Aktivität genauer untersuchen und die möglichen Gründe sehr granular betrachten. So kann ich evtl. feststellen das Mitarbeiter bei einer bestimmten Art von Anfrage sehr viel Zeit in Salesforce verbringen, um Informationen zu sammeln. Hier liegt also viel Potential versteckt, um den gesamten Prozess zu verbessern. In dem z.B. die Informationsbeschaffung erleichtert wird oder evtl. der Anfragetyp optimiert wird, kann dieses Potential genutzt werden. Auch ist Task Mining die ideale Grundlage zur Formulierung von RPA-Lösungen.

Ebenfalls entscheidend für die Zukunftsfähigkeit von Process Mining ist die Möglichkeit, Verknüpfungen zwischen unterschiedlichen Geschäftsprozesse zu erkennen. Häufig sind diese untrennbar miteinander verbunden und der Output eines Prozesses bildet den Input für einen anderen. Mit prozessübergreifenden Multi-Event Logs bietet Celonis die Möglichkeit, genau diese Verbindungen aufzuzeigen. So entsteht ein einheitliches Prozessmodell für das gesamte Unternehmen. Und das unter bestimmten Voraussetzungen auch in nahezu Echtzeit.

Werden die ersten Entwicklungen im Bereich Machine Learning und Task Mining von Celonis weiter ausgebaut, ist Celonis weiterhin auf einem zukunftssicheren Weg. Unternehmen, die vor allem viel Wert auf Enterprise-Readiness und eine intensive Weiterentwicklung legen, dürften mit Celonis auf der sicheren Seite sein.

Preisgestaltung

Die Preisgestaltung der Enterprise Version wird von Celonis nicht transparent kommuniziert. Angeboten werden verschiedene kostenpflichtige Lösungspakete, welche sich aus den Anforderungen eines Projektes ergeben.  Generell stufe ich die Celonis Enterprise Version als Premium Produkt ein. Dies liegt auch daran, weil die Basisausführung der Celonis Enterprise Version bereits sehr umfänglich ist und neben der Software Subscription standardmäßig auch mit Wartung und Support kommt. Zusätzlich steckt mittlerweile sehr viel Entwicklungsarbeit in der Celonis Process Mining Plattform, welche weit über klassische Process Discovery Solutions hinausgeht.  Für kleinere Unternehmen mit begrenztem Budget gibt es daher zwischen der kostenfreien Snap Version und den Basis Paketen der Enterprise Version oft keine Interimslösung.

Fazit

Insgesamt stellt Celonis ein unabhängiges und leistungsstarkes Process Mining Tool in der Cloud bereit. Gehört die Cloud zur Unternehmensstrategie, ist man bei Celonis an der richtigen Adresse. Die „prebuild Process-Connectors“ und die vordefinierten Analysen können ein Process Mining Projekt signifikant beschleunigen und somit die Time-to-Value lukrativ verkürzen. Die Analyse Tools sind leicht bedienbar und schaffen dank integrierter Machine Learning Algorithmen Optimierungspotentiale. Positiv ist auch zu bewerten, dass Celonis ohne speziellen Syntax auskommt und mittelmäßige SQL-Fähigkeiten somit völlig ausreichend sind, um Prozessanalysen vollumfänglich durchzuführen. Diesen vielen positiven Aspekten steht eigentlich nur die hohe Preisgestaltung für die Enterprise Version gegenüber. Ob diese im Einzelfall gerechtfertigt ist, sollte situationsabhängig evaluiert werden. Sicherlich richtet sich Celonis Enterprise in erster Linie an größere Unternehmen, welche komplexe Prozesse mit hohen Datenvolumina analysieren möchte.  Mit Celonis-Snap können jedoch auch kleine Unternehmen und Start-ups einen begrenzten Einblick in dieses gut gelungene Process Mining Tool erhalten.