Posts

Attribution Models in Marketing

Attribution Models

A Business and Statistical Case

INTRODUCTION

A desire to understand the causal effect of campaigns on KPIs

Advertising and marketing costs represent a huge and ever more growing part of the budget of companies. Studies have found out this share is as high as 10% and increases with the size of companies (CMO study by American Marketing Association and Duke University, 2017). Measuring precisely the impact of a specific marketing campaign on the sales of a company is a critical step towards an efficient allocation of this budget. Would the return be higher for an euro spent on a Facebook ad, or should we better spend it on a TV spot? How much should I spend on Twitter ads given the volume of sales this channel is responsible for?

Attribution Models have lately received great attention in Marketing departments to answer these issues. The transition from offline to online marketing methods has indeed permitted the collection of multiple individual data throughout the whole customer journey, and  allowed for the development of user-centric attribution models. In short, Attribution Models use the information provided by Tracking technologies such as Google Analytics or Webtrekk to understand customer journeys from the first click on a Facebook ad to the final purchase and adequately ponderate the different marketing campaigns encountered depending on their responsibility in the final conversion.

Issues on Causal Effects

A key question then becomes: how to declare a channel is responsible for a purchase? In other words, how can we isolate the causal effect or incremental value of a campaign ?

          1. A/B-Tests

One method to estimate the pure impact of a campaign is the design of randomized experiments, wherein a control and treated groups are compared.  A/B tests belong to this broad category of randomized methods. Provided the groups are a priori similar in every aspect except for the treatment received, all subsequent differences may be attributed solely to the treatment. This method is typically used in medical studies to assess the effect of a drug to cure a disease.

Main practical issues regarding Randomized Methods are:

  • Assuring that control and treated groups are really similar before treatment. Uually a random assignment (i.e assuring that on a relevant set of observable variables groups are similar) is realized;
  • Potential spillover-effects, i.e the possibility that the treatment has an impact on the non-treated group as well (Stable unit treatment Value Assumption, or SUTVA in Rubin’s framework);
  • The costs of conducting such an experiment, and especially the costs linked to the deliberate assignment of individuals to a group with potentially lower results;
  • The number of such experiments to design if multiple treatments have to be measured;
  • Difficulties taking into account the interaction effects between campaigns or the effect of spending levels. Indeed, usually A/B tests are led by cutting off temporarily one campaign entirely and measuring the subsequent impact on KPI’s compared to the situation where this campaign is maintained;
  • The dynamical reproduction of experiments if we assume that treatment effects may change over time.

In the marketing context, multiple campaigns must be tested in a dynamical way, and treatment effect is likely to be heterogeneous among customers, leading to practical issues in the lauching of A/B tests to approximate the incremental value of all campaigns. However, sites with a lot of traffic and conversions can highly benefit from A/B testing as it provides a scientific and straightforward way to approximate a causal impact. Leading companies such as Uber, Netflix or Airbnb rely on internal tools for A/B testing automation, which allow them to basically test any decision they are about to make.

References:

Books:

Experiment!: Website conversion rate optimization with A/B and multivariate testing, Colin McFarland, ©2013 | New Riders  

A/B testing: the most powerful way to turn clicks into customers. Dan Siroker, Pete Koomen; Wiley, 2013.

Blogs:

https://eng.uber.com/xp

https://medium.com/airbnb-engineering/growing-our-host-community-with-online-marketing-9b2302299324

Study:

https://cmosurvey.org/wp-content/uploads/sites/15/2018/08/The_CMO_Survey-Results_by_Firm_and_Industry_Characteristics-Aug-2018.pdf

        2. Attribution models

Attribution Models do not demand to create an experimental setting. They take into account existing data and derive insights from the variability of customer journeys. One key difficulty is then to differentiate correlation and causality in the links observed between the exposition to campaigns and purchases. Indeed, selection effects may bias results as exposure to campaigns is usually dependant on user-characteristics and thus may not be necessarily independant from the customer’s baseline conversion probabilities. For example, customers purchasing from a discount price comparison website may be intrinsically different from customers buying from FB ad and this a priori difference may alone explain post-exposure differences in purchasing bahaviours. This intrinsic weakness must be remembered when interpreting Attribution Models results.

                          2.1 General Issues

The main issues regarding the implementation of Attribution Models are linked to

  • Causality and fallacious reasonning, as most models do not take into account the aforementionned selection biases.
  • Their difficult evaluation. Indeed, in almost all attribution models (except for those based on classification, where the accuracy of the model can be computed), the additionnal value brought by the use of a given attribution models cannot be evaluated using existing historical data. This additionnal value can only be approximated by analysing how the implementation of the conclusions of the attribution model have impacted a given KPI.
  • Tracking issues, leading to an uncorrect reconstruction of customer journeys
    • Cross-device journeys: cross-device issue arises from the use of different devices throughout the customer journeys, making it difficult to link datapoints. For example, if a customer searches for a product on his computer but later orders it on his mobile, the AM would then mistakenly consider it an order without prior campaign exposure. Though difficult to measure perfectly, the proportion of cross-device orders can approximate 20-30%.
    • Cookies destruction makes it difficult to track the customer his the whole journey. Both regulations and consumers’ rising concerns about data privacy issues mitigate the reliability and use of cookies.1 – From 2002 on, the EU has enacted directives concerning privacy regulation and the extended use of cookies for commercial targeting purposes, which have highly impacted marketing strategies, such as the ‘Privacy and Electronic Communications Directive’ (2002/58/EC). A research was conducted and found out that the adoption of this ‘Privacy Directive’ had led to 64% decrease in advertising methods compared to the rest of the world (Goldfarb et Tucker (2011)). The effect was stronger for generalized sites (Yahoo) than for specialized sites.2 – Users have grown more and more conscious of data privacy issues and have adopted protective measures concerning data privacy, such as automatic destruction of cookies after a session is ended, or simply giving away less personnal information (Goldfarb et Tucker (2012) ) .Valuable user information may be lost, though tracking technologies evolution have permitted to maintain tracking by other means. This issue may be particularly important in countries highly concerned with data privacy issues such as Germany.
    • Offline/Online bridge: an Attribution Model should take into account all campaigns to draw valuable insights. However, the exposure to offline campaigns (TV, newspapers) are difficult to track at the user level. One idea to tackle this issue would be to estimate the proportion of conversions led by offline campaigns through AB testing and deduce this proportion from the credit assigned to the online campaigns accounted for in the Attribution Model.
    • Touch point information available: clicks are easy to follow but irrelevant to take into account the influence of purely visual campaigns such as display ads or video.

                          2.2 Today’s main practices

Two main families of Attribution Models exist:

  • Rule-Based Attribution Models, which have been used for in the last decade but from which companies are gradualy switching.

Attribution depends on the individual journeys that have led to a purchase and is solely based on the rank of the campaign in the journey. Some models focus on a single touch points (First Click, Last Click) while others account for multi-touch journeys (Bathtube, Linear). It can be calculated at the customer level and thus doesn’t require large amounts of data points. We can distinguish two sub-groups of rule-based Attribution Models:

  • One Touch Attribution Models attribute all credit to a single touch point. The First-Click model attributes all credit for a converion to the first touch point of the customer journey; last touch attributes all credit to the last campaign.
  • Multi-touch Rule-Based Attribution Models incorporate information on the whole customer journey are thus an improvement compared to one touch models. To this family belong Linear model where credit is split equally between all channels, Bathtube model where 40% of credit is given to first and last clicks and the remaining 20% is distributed equally between the middle channels, or time-decay models where credit assigned to a click diminishes as the time between the click and the order increases..

The main advantages of rule-based models is their simplicity and cost effectiveness. The main problems are:

– They are a priori known and can thus lead to optimization strategies from competitors
– They do not take into account aggregate intelligence on customer journeys and actual incremental values.
– They tend to bias (depending on the model chosen) channels that are over-represented at the beggining or end of the funnel, according to theoretical assumptions that have no observationnal back-ups.

  • Data-Driven Attribution Models

These models take into account the weaknesses of rule-based models and make a relevant use of available data. Being data-driven, following attribution models cannot be computed using single user level data. On the contrary values are calculated through data aggregation and thus require a certain volume of customer journey information.

References:

https://dspace.mit.edu/handle/1721.1/64920

 

        3. Data-Driven Attribution Models in practice

                          3.1 Issues

Several issues arise in the computation of campaigns individual impact on a given KPI within a data-driven model.

  • Selection biases: Exposure to certain types of advertisement is usually highly correlated to non-observable variables which are in turn correlated to consumption practices. Differences in the behaviour of users exposed to different campaigns may thus only be driven by core differences in conversion probabilities between groups whether than by the campaign effect.
  • Complementarity: it may be that campaigns A and B only have an effect when combined, so that measuring their individual impact would lead to misleading conclusions. The model could then try to assess the effect of combinations of campaigns on top of the effect of individual campaigns. As the number of possible non-ordered combinations of k campaigns is 2k, it becomes clear that inclusing all possible combinations would however be time-consuming.
  • Order-sensitivity: The effect of a campaign A may depend on the place where it appears in the customer journey, meaning the rank of a campaign and not merely its presence could be accounted for in the model.
  • Relative Order-sensitivity: it may be that campaigns A and B only have an effect when one is exposed to campaign A before campaign B. If so, it could be useful to assess the effect of given combinations of campaigns as well. And this for all campaigns, leading to tremendous numbers of possible combinations.
  • All previous phenomenon may be present, increasing even more the potential complexity of a comprehensive Attribution Model. The number of all possible ordered combination of k campaigns is indeed :

 

                          3.2 Main models

                                  A) Logistic Regression and Classification models

If non converting journeys are available, Attribition Model can be shaped as a simple classification issue. Campaign types or campaigns combination and volume of campaign types can be included in the model along with customer or time variables. As we are interested in inference (on campaigns effect) whether than prediction, a parametric model should be used, such as Logistic Regression. Non paramatric models such as Random Forests or Neural Networks can also be used though the interpretation of campaigns value would be more difficult to derive from the model results.

A common pitfall is the usual issue of spurious correlations on one hand and the correct interpretation of coefficients in business terms.

An advantage if the possibility to evaluate the relevance of the model using common model validation methods to evaluate its predictive power (validation set \ AUC \pseudo R squared).

                                  B) Shapley Value

Theory

The Shapley Value is based on a Game Theory framework and is named after its creator, the Nobel Price Laureate Lloyd Shapley. Initially meant to calculate the marginal contribution of players in cooperative games, the model has received much attention in research and industry and has lately been applied to marketing issues. This model is typically used by Google Adords and other ad bidding vendors. Campaigns or marketing channels are in this model seen as compementary players looking forward to increasing a given KPI.
Contrarily to Logistic Regressions, it is a non-parametric model. Contrarily to Markov Chains, all results are built using existing journeys, and not simulated ones.

Channels are considered to enter the game sequentially under a certain joining order. Shapley value try to The Shapley value of channel i is the weighted sum of the marginal values that channel i adds to all possible coalitions that don’t contain channel i.
In other words, the main logic is to analyse the difference of gains when a channel i is added after a coalition Ck of k channels, k<=n. We then sum all the marginal contributions over all possible ordered combination Ck of all campaigns excluding i, with k<=n-1.

Subsets framework

A first an most usual way to compute the Shapley Vaue is to consider that when a channel enters coalition, its additionnal value is the same irrelevant of the order in which previous channels have appeared. In other words, journeys (A>B>C) and (B>A>C) trigger the same gains.
Shapley value is computed as the gains associated to adding a channel i to a subset of channels, weighted by the number of (ordered) sequences that the (unordered) subset represents, summed up on all possible subsets of the total set of campaigns where the channel i is not present.
The Shapley value of the channel 𝑥𝑗 is then:

where |S| is the number of campaigns of a coalition S and the sum extends over all subsets S that do not not contain channel j. 𝜈(𝑆)  is the value of the coalition S and 𝜈(𝑆 ∪ {𝑥𝑗})  the value of the coalition formed by adding 𝑥𝑗 to coalition S. 𝜈(𝑆 ∪ {𝑥𝑗}) − 𝜈(𝑆) is thus the marginal contribution of channel 𝑥𝑗 to the coalition S.

The formula can be rewritten and understood as:

This method is convenient when data on the gains of on all possible permutations of all unordered k subsets of the n campaigns are available. It is also more convenient if the order of campaigns prior to the introduction of a campaign is thought to have no impact.

Ordered sequences

Let us define 𝜈((A>B)) as the value of the sequence A then B. What is we let 𝜈((A>B)) be different from 𝜈((B>A)) ?
This time we would need to sum over all possible permutation of the S campaigns present before  𝑥𝑗 and the N-(S+1) campaigns after 𝑥𝑗. Doing so we will sum over all possible orderings (i.e all permutations of the n campaigns of the grand coalition containing all campaigns) and we can remove the permutation coefficient s!(p-s+1)!.

This method is convenient when the order of channels prior to and after the introduction of another channel is assumed to have an impact. It is also necessary to possess data for all possible permutations of all k subsets of the n campaigns, and not only on all (unordered) k-subsets of the n campaigns, k<=n. In other words, one must know the gains of A, B, C, A>B, B>A, etc. to compute the Shapley Value.

Differences between the two approaches

We simulate an ordered case where the value for each ordered sequence k for k<=3 is known. We compare it to the usual Shapley value calculated based on known gains of unordered subsets of campaigns. So as to compare relevant values, we have built the gains matrix so that the gains of a subset A, B i.e  𝜈({B,A}) is the average of the gains of ordered sequences made up with A and B (assuming the number of journeys where A>B equals the number of journeys where B>A, we have 𝜈({B,A})=0.5( 𝜈((A>B)) + 𝜈((B>A)) ). We let the value of the grand coalition be different depending on the order of campaigns-keeping the constraints that it averages to the value used for the unordered case.

Note: mvA refers to the marginal value of A in a given sequence.
With traditionnal unordered coalitions:

With ordered sequences used to compute the marginal values:

 

We can see that the two approaches yield very different results. In the unordered case, the Shapley Value campaign C is the highest, culminating at 20, while A and B have the same Shapley Value mvA=mvB=15. In the ordered case, campaign A has the highest Shapley Value and all campaigns have different Shapley Values.

This example illustrates the inherent differences between the set and sequences approach to Shapley values. Real life data is more likely to resemble the ordered case as conversion probabilities may for any given set of campaigns be influenced by the order through which the campaigns appear.

Advantages

Shapley value has become popular in allocation problems in cooperative games because it is the unique allocation which satisfies different axioms:

  • Efficiency: Shaple Values of all channels add up to the total gains (here, orders) observed.
  • Symmetry: if channels A and B bring the same contribution to any coalition of campaigns, then their Shapley Value i sthe same
  • Null player: if a channel brings no additionnal gains to all coalitions, then its Shapley Value is zero
  • Strong monotony: the Shapley Value of a player increases weakly if all its marginal contributions increase weakly

These properties make the Shapley Value close to what we intuitively define as a fair attribution.

Issues

  • The Shapley Value is based on combinatory mathematics, and the number of possible coalitions and ordered sequences becomes huge when the number of campaigns increases.
  • If unordered, the Shapley Value assumes the contribution of campaign A is the same if followed by campaign B or by C.
  • If ordered, the number of combinations for which data must be available and sufficient is huge.
  • Channels rarely present or present in long journeys will be played down.
  • Generally, gains are supposed to grow with the number of players in the game. However, it is plausible that in the marketing context a journey with a high number of channels will not necessarily bring more orders than a journey with less channels involved.

References:

R package: GameTheoryAllocation

Article:
Zhao & al, 2018 “Shapley Value Methods for Attribution Modeling in Online Advertising “
https://link.springer.com/content/pdf/10.1007/s13278-017-0480-z.pdf
Courses: https://www.lamsade.dauphine.fr/~airiau/Teaching/CoopGames/2011/coopgames-7%5b8up%5d.pdf
Blogs: https://towardsdatascience.com/one-feature-attribution-method-to-supposedly-rule-them-all-shapley-values-f3e04534983d

                                  B) Markov Chains

Markov Chains are used to model random processes, i.e events that occur in a sequential manner and in such a way that the probability to move to a certain state only depends on the past steps. The number of previous steps that are taken into account to model the transition probability is called the memory parameter of the sequence, and for the model to have a solution must be comprised between 0 and 4. A Markov Chain process is thus defined entirely by its Transition Matrix and its initial vector (i.e the starting point of the process).

Markov Chains are applied in many scientific fields. Typically, they are used in weather forecasting, with the sequence of Sunny and Rainy days following a Markov Process of memory parameter 0, so that for each given day the probability that the next day will be rainy or sunny only depends on the weather of the current day. Other applications can be found in sociology to understand the dynamics of social classes intergenerational reproduction. To get more both mathematical and applied illustration, I recommend the reading of this course.

In the marketing context, Markov Chains are an interesting way to model the conversion funnel. To go from the from the Markov Model to the Attribution logic, we calculate the Removal Effect of each channel, i.e the difference in conversions that happen if the channel is removed. Please read below for an introduction to the methodology.

The first step in a Markov Chains Attribution Model is to build the transition matrix that captures the transition probabilities between the campaigns accross existing customer journeys. This Matrix is to be read as a “From state A to state B” table, from the left to the right. A first difficulty is finding the right memory parameter to use. A large memory parameter would allow to take more into account interraction effects within the conversion funnel but would lead to increased computationnal time, a non-readable transition matrix, and be more sensitive to noisy data. Please note that this transition matrix provides useful information on the conversion funnel and on the relationships between campaigns and can be used as such as an analytical tool. I suggest the clear and easily R code which can be found here or here.

Here is an illustration of a Markov Chain with memory Parameter of 0: the probability to go to a certain campaign B in the next step only depend on the campaign we are currently at:

The associated Transition Matrix is then (with null probabilities left as Blank):

The second step is  to compute the actual responsibility of a channel in total conversions. As mentionned above, the main philosophy to do so is to calculate the Removal Effect of each channel, i.e the changes in the number of conversions when a channel is entirely removed. All customer journeys which went through this channel are settled out to be unsuccessful. This calculation is done by applying the transition matrix with and without the removed channels to an initial vector that contains the number of desired simulations.

Building on our current example, we can then settle an initial vector with the desired number of simulations, e.g 10 000:

 

It is possible at this stage to add a constraint on the maximum number of times the matrix is applied to the data, i.e on the maximal number of campaigns a simulated journey is allowed to have.

Advantages

  • The dynamic journey is taken into account, as well as the transition between two states. The funnel is not assumed to be linear.
  • It is possile to build a conversion graph that maps the customer journey provides valuable insights.
  • It is possible to evaluate partly the accuracy of the Attribution Model based on Markov Chains. It is for example possible to see how well the transition matrix help predict the future by analysing the number of correct predictions at any given step over all sequences.

Disadvantages

  • It can be somewhat difficult to set the memory parameter. Complementarity effects between channels are not well taken into account if the memory is low, but a parameter too high will lead to over-sensitivity to noise in the data and be difficult to implement if customer journeys tend to have a number of campaigns below this memory parameter.
  • Long journeys with different channels involved will be overweighted, as they will count many times in the Removal Effect.  For example, if there are n-1 channels in the customer journey, this journey will be considered as failure for the n-1 channel-RE. If the volume effects (i.e the impact of the overall number of channels in a journey, irrelevant from their type° are important then results may be biased.

References:

R package: ChannelAttribution

Git:

https://github.com/MatCyt/Markov-Chain/blob/master/README.md

Course:

https://www.ssc.wisc.edu/~jmontgom/markovchains.pdf

Article:

“Mapping the Customer Journey: A Graph-Based Framework for Online Attribution Modeling”; Anderl, Eva and Becker, Ingo and Wangenheim, Florian V. and Schumann, Jan Hendrik, 2014. Available at SSRN: https://ssrn.com/abstract=2343077 or http://dx.doi.org/10.2139/ssrn.2343077

“Media Exposure through the Funnel: A Model of Multi-Stage Attribution”, Abhishek & al, 2012

“Multichannel Marketing Attribution Using Markov Chains”, Kakalejčík, L., Bucko, J., Resende, P.A.A. and Ferencova, M. Journal of Applied Management and Investments, Vol. 7 No. 1, pp. 49-60.  2018

Blogs:

https://analyzecore.com/2016/08/03/attribution-model-r-part-1

https://analyzecore.com/2016/08/03/attribution-model-r-part-2

                          3.3 To go further: Tackling selection biases with Quasi-Experiments

Exposure to certain types of advertisement is usually highly correlated to non-observable variables. Differences in the behaviour of users exposed to different campaigns may thus only be driven by core differences in converison probabilities between groups whether than by the campaign effect. These potential selection effects may bias the results obtained using historical data.

Quasi-Experiments can help correct this selection effect while still using available observationnal data.  These methods recreate the settings on a randomized setting. The goal is to come as close as possible to the ideal of comparing two populations that are identical in all respects except for the advertising exposure. However, populations might still differ with respect to some unobserved characteristics.

Common quasi-experimental methods used for instance in Public Policy Evaluation are:

  • Discontinuity Regressions
  • Matching Methods, such as Exact Matching,  Propensity-score matching or k-nearest neighbourghs.

References:

Article:

“Towards a digital Attribution Model: Measuring the impact of display advertising on online consumer behaviour”, Anindya Ghose & al, MIS Quarterly Vol. 40 No. 4, pp. 1-XX, 2016

https://pdfs.semanticscholar.org/4fa6/1c53f281fa63a9f0617fbd794d54911a2f84.pdf

        4. First Steps towards a Practical Implementation

Identify key points of interests

  • Identify the nature of touchpoints available: is the data based on clicks? If so, is there a way to complement the data with A/B tests to measure the influence of ads without clicks (display, video) ? For example, what happens to sales when display campaign is removed? Analysing this multiplier effect would give the overall responsibility of display on sales, to be deduced from current attribution values given to click-based channels. More interestingly, what is the impact of the removal of display campaign on the occurences of click-based campaigns ? This would give us an idea of the impact of display ads on the exposure to each other campaigns, which would help correct the attribution values more precisely at the campaign level.
  • Define the KPI to track. From a pure Marketing perspective, looking at purchases may be sufficient, but from a financial perspective looking at profits, though a bit more difficult to compute, may drive more interesting results.
  • Define a customer journey. It may seem obvious, but the notion needs to be clarified at first. Would it be defined by a time limit? If so, which one? Does it end when a conversion is observed? For example, if a customer makes 2 purchases, would the campaigns he’s been exposed to before the first order still be accounted for in the second order? If so, with a time decay?
  • Define the research framework: are we interested only in customer journeys which have led to conversions or in all journeys? Keep in mind that successful customer journeys are a non-representative sample of customer journeys. Models built on the analysis of biased samples may be conservative. Take an extreme example: 80% of customers who see campaign A buy the product, VS 1% for campaign B. However, campaign B exposure is great and 100 Million people see it VS only 1M for campaign A. An Attribution Model based on successful journeys will give higher credit to campaign B which is an auguable conclusion. Taking into account costs per campaign (in the case where costs are calculated by clicks) may of course tackle this issue partly, as campaign A could then exhibit higher returns, but a serious fallacious reasonning is at stake here.

Analyse the typical customer journey    

  • Performing a duration analysis on the data may help you improve the definition of the customer journey to be used by your organization. After which days are converison probabilities null? Should we consider the effect of campaigns disappears after x days without orders? For example, if 99% of orders are placed in the 30 days following a first click, it might be interesting to define the customer journey as a 30 days time frame following the first oder.
  • Look at the distribution of the number of campaigns in a typical journey. If you choose to calculate the effect of campaigns interraction in your Attribution Model, it may indeed help you determine the maximum number of campaigns to be included in a combination. Indeed, you may not need to assess the impact of channel combinations with above than 4 different channels if 95% of orders are placed after less then 4 campaigns.
  • Transition matrixes: what if a campaign A systematically leads to a campaign B? What happens if we remove A or B? These insights would give clues to ask precise questions for a latter AB test, for example to find out if there is complementarity between channels A and B – (implying none should be removed) or mere substitution (implying one can be given up).
  • If conversion rates are available: it can be interesting to perform a survival analysis i.e to analyse the likelihood of conversion based on duration since first click. This could help us excluse potential outliers or individuals who have very low conversion probabilities.

Summary

Attribution is a complex topic which will probably never be definitively solved. Indeed, a main issue is the difficulty, or even impossibility, to evaluate precisely the accuracy of the attribution model that we’ve built. Attribution Models should be seen as a good yet always improvable approximation of the incremental values of campaigns, and be presented with their intrinsinc limits and biases.

MeetUp: Classifying Text Data with Embeddings // How to become a data scientist

Welcome to our 3rd DATANOMIQ Data Science MeetUp, in Berlin!

April 3rd: 6 pm until 8 pm.

Click the link to participate:
https://www.meetup.com/de-DE/DATANOMIQ-Data-Science-Berlin/events/257098910/

Today’s topic is all about Classifying Text Data with Embeddings, presented by Artur Zeitler. Artur is a data scientist at DATANOMIQ.
Make sure to come early to grab a voucher for a free drink.

Agenda:
doors open at 6 pm (drinks)

PART ONE (6:30 pm):
– The advantages of word embeddings compared to Bag-of-Words
– The differences between the popular embeddings word2vec and GloVe
– How doc2vec can numerically represent whole documents
– How to apply doc2vec in Tensorflow to classify documents
– open questions (7pm)
———————————————————————-
15 min break
———————————————————————-

PART TWO (7:15 pm):
– How to start a career as a data scientist
(presented by Benjamin Aunkofer. Benjamin is Chief Data Scientist at DATANOMIQ.)
– Open questions

FREE ENTRY

Click the link to participate:
https://www.meetup.com/de-DE/DATANOMIQ-Data-Science-Berlin/events/257098910/

 

Next MeetUp:
May 8: Imbalanced datasets: Dealing with the minority class
https://www.meetup.com/de-DE/DATANOMIQ-Data-Science-Berlin/events/259723993/

A Gentle Introduction to Precision and Recall.

The idea of this blog is to give an intuitive understanding of Precision and Recall for a binary classification problem. I will shy away from explaining it in a textbook way but rather will try to give an intuition. Nevertheless, let me write the textbook formula first:

The problem with this nomenclature is that despite being correct, it can be a bit confusing, especially for beginners. For example ‘False Positives’ could be understood from a classifier point of view or from a population point of view.

Visualizing with an example

Let’s suppose we have a classifier to differentiate jeans from a T-shirts in a lot of cloths. This lot has 100 pieces altogether with 70 jeans and 30 T-shirts. Let us see this visually. Until this point, we just have a collection of clothes and have no classifier.

We already know that altogether we truly have 70 Jeans and 30 T-shirts.

Now let’s run the classifier to identify the jeans from T-shirts. We can assume the result of the classifier is following (number inside the box is the result of classifier):

We see that out of 70 jeans the classifier identifies 63 correctly as jeans and the remaining 7 as non-Jeans. Out of 30 T-shirts, the classifier identifies 11 falsely as jeans the remaining 19 correctly as non-Jeans.

So Recall is nothing but the proportion of identified jeans out of total jeans, which is

Recall = 63 / 70

Precision is the true jeans identified out of the total number of classified jeans. Which is:

Precision = 63 / (63+11)

Hence we see, in a way Recall has to do with the ability of classifier to deal with jeans and precision has to do with ability to deal with both Jeans and Non-Jeans.

This seems to provide better intuition than the textbook formula.

Diving Deeper with another example

Let us go through one more example to cement the idea. Let’s imagine there is a village which has a notoriously high number of criminals. A special cop arrives to tackle the law and order situation. He interviews every resident and locks some residents based on hunches.

If there are still many criminals roaming on the street the recall is bad, as recall deals with the ability to deal with the quantity which classifier is supposed to find (in this case criminals).

If there are too many innocents rotting in jail the precision is bad. As precision has also to do with the ability to deal with ‘others‘ that is not the quantity which the classifier is supposed to find (in this case these are the innocents).

Now we see, we don’t want too many criminals roaming on the street nor do we want many innocents rotting in the jail. Hence we need both recall and precision to be high or in other words, their mean to be high. But this cannot be arithmetic mean. Let’s see why using an example.

If for a village of 2000 residents there are 100 criminals. And if the cop straight away locks all 2000 residents, the confusion matrix looks like this:

 

Recall= 100/ (100+0) = 1

Precision = 100/ (100+1900) = 0.05

Arithmetic mean for Precision and Recall = (1+.05)/2 = 0.525

This would look like a pretty good classifier even though we know that in reality it’s a bad classifier (or a bad cop who just locks up every person he meets). It can be shown that the same happens in reverse. If the cop does not lock up anyone, the arithmetic mean does not show the true picture again.

That’s why we use harmonic mean. We call it F1 Score and it is calculated as follows: (2 * 1 * 0.05) / (1 + 0.05) = 0.0952

Now, this looks like a more realistic score. So, the performance of a classifier can be judged with a harmonic mean between precision and recall.

Let’s try to understand one more thing.

Often, classifiers work by returning probabilities of positives and negatives. One way to turn them into a confusion matrix is to use a threshold of 0.5. This means that if the probability of being positive is more than 0.5, we consider the case as positive (in our case a criminal). Otherwise, it is a negative.

But there might be cases where we want our recall to be very high. For example, if there is a classifier for identifying Ebola. We do not want any of the cases to be missed because otherwise we are risking an outbreak of the decease with disastrous consequences.

In this case, the threshold needs to be kept really low (maybe near .1 or smaller) so that we raise a flag for every case that has at least 10 % probability and get this person retested. This is an important measure in order to prevent an outbreak, despite the fact that there are a lot of false cases that needs to be rechecked.

There might be other cases where there are many false alarms (maybe fraud transaction in banks) which may be of low risk and it would be expensive to investigate all those cases. In those case, we might want to have a threshold higher than 0.5.

This gives us a taste of things to come. A classifiers efficiency can be plotted for different thresholds which gives us something called a ROC curve. But let’s save that for another post.

How is automation changing data science and machine learning?

We have come a long way since the introduction of data science and machine learning. The recent study has found that the volume of business data doubles in less than 14 months. Today, the collection of data is no longer a problem, but the filtration, analysis, and maintenance of relevant information is a bigger issue.

We need to hire data science professionals, and they demand over $100k annually. Paying that sort of money for a professional is not feasible for every single organization, especially small and middle-sized companies. Google recently announced that it is going to make machine learning technology possible for every business.

The access to machine learning technology is now possible, even for small businesses due to automation. Google, Microsoft, and other companies have come up with automated machine learning tools that enable small businesses to use machine learning technology to enhance their business performance and profit.

Image Source: Google Cloud

With that said, the world still needs a lot of machine learning professionals. Many machine learning professionals prefer Python for machine learning due to its features and a wide range of libraries.

According to the Gartner report, around 40% of data science tasks will be automated by 2020. The data science tools can automate some parts of data science processes, but it is not complete automation.

With that said, it has been helping a lot to accelerate the tasks. We still need data science professionals to deal with real-world problems. The algorithms are not yet able to handle messy data. The significant chunk of data science professionals often prefers performing with data science with Python for sophisticated tasks.

Automation in Data Science

Let me show you the figure right at the beginning before moving forward.

Image Source: Wikipedia

If I had to use only one word to describe the entire data science process, I would use the word “headache.” According to the recent report, the median salary of data scientists easily surpasses $100k annually. The pay will be higher in the time to come.

One needs to pay a lot of money and invest a lot of time to get insights from the collected data. The data scientists need to spend almost 50-60% of their time in data processing and the rest of their time in modeling and deployment.

The cloud platforms like Amazon Web Services, Google, Microsoft Azure, and so on make the job more comfortable, but there is still a lot of work to maintain and extract useful insights from the collected data.

The data science process has lots of inefficiencies. At first, they need to spend over 50% of their total time on processing messy real-world data. After that, there could be a need to customize models, according to specific problems.

The significant contribution of automation is making a significant portion of data processing parts automated. Secondly, the automated platforms can make tracking of various models easier from multiple parameters. The time needed to launch the algorithm is minimal.

One example of an extensive tool to handle a data science project is Alteryx. IT has come up with powerful automated solutions that can drastically reduce the data processing and model development time for smoothening the entire data science workflow. The data science platform, Alteryx, is so amazing that its share price doubled in a span of little more than a year.

Some other great tools that can help you in data science automation are Rapidminer, H20.ai, KNIME, and so on. However, the lack of skilled data scientists can create a problem despite these tools. It is where the role of automated machine learning pops in.

How is Machine Learning Transformed with the entrance of Automation?

The traditional machine learning process was too complicated. One requires to have a lot of expensive machine learning professionals working for months to come up with models to process machine learning tasks.

Image Source: Medium

To make traditional machine learning work, one needs to gather data, standardize data, process features, create and train the machine learning model from problems, validate the models, and deploy the models at last.

You must have heard of how machine learning is only for corporations in the past. But, that has drastically changed in recent time, and it is all due to automation. Keep in mind that the above machine learning model is a simple one. There is a lot of extra works for complicated models. Even for the simple ones, you need to spend a lot of time and money, which makes it impossible for small and medium companies.

The automation in machine learning is all about automating the entire process to make machine learning easier. The only thing you need to do is feed data to the system (not a massive volume of data). You do not need even to cross the three-figure number of images to continue with automated machine learning platforms.

Microsoft has its automl platform along with Google. Other automl platforms can do the trick for you. Using those platforms do not cost you an arm and a leg. If you check out the price, you will be surprised.

There is no need for you to create or deploy models or even test the models. The algorithm will do the job for you. It takes examples and models of historical models to process the data and use a machine learning algorithm.

Even non-statistician can implement machine learning technology with limited data, thanks to automation in machine learning. You can make use of predictive analytics and can get easy solutions for simple prediction problems without scratching your head. Numerous libraries can assist you in the automated generation of machine learning pipelines.

How are the jobs of data scientists simplified by the introduction of automation in machine learning and data science?

It is true that the introduction of automation has drastically reduced the time for completing the tasks for data scientists. They no longer have to spend their valuable time in time-consuming, monotonous works that are necessary but do not provide a lot of value.

However, the need for skilled data scientists still exist, and it will always be there in the time to come. There are challenging works for data scientists that we cannot replace with machines, such as listening to clients, figuring out the root cause of business issues, development and selection of the right solution for the specific business problem.

Just like in other types of jobs, the advancement of automation technologies will modify the tasks that data scientists need to perform. They will be able to allocate more time on things that matter rather than monotonous tasks.

Final Verdict

The automation of machine learning and data science are in the beginning stage. However, they are already making a massive impact on the business world. The huge corporations are investing in Big Data and Machine Learning technologies. We can expect a considerable improvement in these technologies shortly.

Sooner, the competitive advantage of a business will depend on how well they can use the technologies, instead of access to machine learning or Big Data technologies.  I hope this article was valuable to you. If you want to add something or express your thoughts, feel free to leave a comment. I will gladly read and reply to your comment.

A common trap when it comes to sampling from a population that intrinsically includes outliers

I will discuss a common fallacy concerning the conclusions drawn from calculating a sample mean and a sample standard deviation and more importantly how to avoid it.

Suppose you draw a random sample x_1, x_2, … x_N of size N and compute the ordinary (arithmetic) sample mean  x_m and a sample standard deviation sd from it.  Now if (and only if) the (true) population mean µ (first moment) and population variance (second moment) obtained from the actual underlying PDF  are finite, the numbers x_m and sd make the usual sense otherwise they are misleading as will be shown by an example.

By the way: The common correlation coefficient will also be undefined (or in practice always point to zero) in the presence of infinite population variances. Hopefully I will create an article discussing this related fallacy in the near future where a suitable generalization to Lévy-stable variables will be proposed.

 Drawing a random sample from a heavy tailed distribution and discussing certain measures

As an example suppose you have a one dimensional random walker whose step length is distributed by a symmetric standard Cauchy distribution (Lorentz-profile) with heavy tails, i.e. an alpha-stable distribution with alpha being equal to one. The PDF of an individual independent step is given by p(x) = \frac{\pi^{-1}}{(1 + x^2)} , thus neither the first nor the second moment exist whereby the first exists and vanishes at least in the sense of a principal value due to symmetry.

Still let us generate N = 3000 (pseudo) standard Cauchy random numbers in R* to analyze the behavior of their sample mean and standard deviation sd as a function of the reduced sample size n \leq N.

*The R-code is shown at the end of the article.

Here are the piecewise sample mean (in blue) and standard deviation (in red) for the mentioned Cauchy sampling. We see that both the sample mean and sd include jumps and do not converge.

Especially the mean deviates relatively largely from zero even after 3000 observations. The sample sd has no target due to the population variance being infinite.

If the data is new and no prior distribution is known, computing the sample mean and sd will be misleading. Astonishingly enough the sample mean itself will have the (formally exact) same distribution as the single step length p(x). This means that the sample mean is also standard Cauchy distributed implying that with a different Cauchy sample one could have easily observed different sample means far of the presented values in blue.

What sense does it make to present the usual interval x_m \pm sd / \sqrt{N} in such a case? What to do?

The sample median, median absolute difference (mad) and Inter-Quantile-Range (IQR) are more appropriate to describe such a data set including outliers intrinsically. To make this plausible I present the following plot, whereby the median is shown in black, the mad in green and the IQR in orange.

This example shows that the median, mad and IQR converge quickly against their assumed values and contain no major jumps. These quantities do an obviously better job in describing the sample. Even in the presence of outliers they remain robust, whereby the mad converges more quickly than the IQR. Note that a standard Cauchy sample will contain half of its sample in the interval median \pm mad meaning that the IQR is twice the mad.

Drawing a random sample from a PDF that has finite moments

Just for comparison I also show the above quantities for a standard normal (pseudo) sample labeled with the same color as before as a counter example. In this case not only do both the sample mean and median but also the sd and mad converge towards their expected values (see plot below). Here all the quantities describe the data set properly and there is no trap since there are no intrinsic outliers. The sample mean itself follows a standard normal, so that the sd in deed makes sense and one could calculate a standard error \frac{sd}{\sqrt{N}} from it to present the usual stochastic confidence intervals for the sample mean.

A careful observation shows that in contrast to the Cauchy case here the sampled mean and sd converge more quickly than the sample median and the IQR. However still the sampled mad performs about as well as the sd. Again the mad is twice the IQR.

And here are the graphs of the prementioned quantities for a pseudo normal sample:

The take-home-message:

Just be careful when you observe outliers and calculate sample quantities right away, you might miss something. At best one carefully observes how the relevant quantities change with sample size as demonstrated in this article.

Such curves should become of broader interest in order to improve transparency in the Data Science process and reduce fallacies as well.

Thank you for reading.

P.S.: Feel free to play with the set random seed in the R-code below and observe how other quantities behave with rising sample size. Of course you can also try different PDFs at the beginning of the code. You can employ a Cauchy, Gaussian, uniform, exponential or Holtsmark (pseudo) random sample.

 

QUIZ: Which one of the recently mentioned random samples contains a trap** and why?

**in the context of this article

 

R-code used to generate the data and for producing plots:

 

 

Business Intelligence Organizations

I am often asked how the Business Intelligence department should be set up and how it should interact and collaborate with other departments. First and foremost: There is no magic recipe here, but every company must find the right organization for itself.

Before we can talk about organization of BI, we need to have a clear definition of roles for team members within a BI department.

A Data Engineer (also Database Developer) uses databases to save structured, semi-structured and unstructured data. He or she is responsible for data cleaning, data availability, data models and also for the database performance. Furthermore, a good Data Engineer has at least basic knowledge about data security and data privacy. A Data Engineer uses SQL and NoSQL-Technologies.

A Data Analyst (also BI Analyst or BI Consultant) uses the data delivered by the Data Engineer to create or adjust data models and implementing business logic in those data models and BI dashboards. He or she needs to understand the needs of the business. This job requires good communication and consulting skills as well as good developing skills in SQL and BI Tools such like MS Power BI, Tableau or Qlik.

A Business Analyst (also Business Data Analyst) is a person form any business department who has basic knowledge in data analysis. He or she has good knowledge in MS Excel and at least basic knowledge in data analysis and BI Tools. A Business Analyst will not create data models in databases but uses existing data models to create dashboards or to adjust existing data analysis applications. Good Business Analyst have SQL Skills.

A Data Scientist is a Data Analyst with extended skills in statistics and machine learning. He or she can use very specific tools and analytical methods for finding pattern in unknow or big data (Data Mining) or to predict events based on pattern calculated by using historized data (Predictive Analytics). Data Scientists work mostly with Python or R programming.

Organization Type 1 – Central Approach (Data Lab)

The first type of organization is the data lab approach. This organization form is easy to manage because it’s focused and therefore clear in terms of budgeting. The data delivery is done centrally by experts and their method and technology knowledge. Consequently, the quality expectation of data delivery and data analysis as well as the whole development process is highest here. Also the data governance is simple and the responsibilities clearly adjustable. Not to be underestimated is the aspect of recruiting, because new employees and qualified applicants like to join a central team of experts.

However, this form of organization requires that the company has the right working attitude, especially in the business intelligence department. A centralized business intelligence department acts as a shared service. Accordingly, customer-oriented thinking becomes a prerequisite for the company’s success – and customers here are the other departments that need access to the capacities of those centralized data experts. Communication boundaries must be overcome and ways of simple and effective communication must be found.

Organization Type 2 – Stakeholder Focus Approach

Other companies want to shift more responsibility for data governance, and especially data use and analytics, to those departments where data plays a key role right now. A central business intelligence department manages its own projects, which have a meaning for the entire company. The specialist departments, which have a special need for data analysis, have their own data experts who carry out critical projects for the specialist department. The central Business Intelligence department does not only provide the technical delivery of data, but also through methodical consulting. Although most of the responsibility lies with the Business Intelligence department, some other data-focused departments are at least co-responsible.

The advantage is obvious: There are special data experts who work deeper in the actual departments and feel more connected and responsible to them. The technical-business focus lies on pain points of the company.

However, this form of Ogranization also has decisive disadvantages: The danger of developing isolated solutions that are so special in some specific areas that they will not really work company-wide increases. Typically the company has to deal with asymmetrical growth of data analytics
know-how. Managing data governance is more complex and recruitment is becoming more difficult as the business intelligence department is weakened and smaller, and data professionals for other departments need to have more business focus, which means they are looking for more specialized profiles.

Organization Type 3 – Decentral Approach

Some companies are also taking a more extreme approach in the other direction. The Business Intelligence department now has only Data Engineers building and maintaining the data warehouse or data lake. As a result, the central department only provides data; it is used and analyzed in all other departments, specifically for the respective applications.

The advantage lies in the personal responsibility of the respective departments as „pain points“ of the company are in focus in belief that business departments know their problems and solutions better than any other department does. Highly specialized data experts can understand colleagues of their own department well and there is no no shared service mindset neccessary, except for the data delivery.

Of course, this organizational form has clear disadvantages since many isolated solutions are unavoidable and the development process of each data-driven solution will be inefficient. These insular solutions may work with luck for your own department, but not for the whole company. There is no one single source of truth. The recruiting process is more difficult as it requires more specialized data experts with more business background. We have to expect an asymmetrical growth of data analytics know-how and a difficult data governance.

 

Fehler-Rückführung mit der Backpropagation

Dies ist Artikel 4 von 6 der Artikelserie –Einstieg in Deep Learning.

Das Gradienten(abstiegs)verfahren ist der Schlüssel zum Training einzelner Neuronen bzw. deren Gewichtungen zu den Neuronen der vorherigen Schicht. Wer dieses Prinzip verstanden hat, hat bereits die halbe Miete zum Verständnis des Trainings von künstlichen neuronalen Netzen.

Der Gradientenabstieg wird häufig fälschlicherweise mit der Backpropagation gleichgesetzt, jedoch ist das nicht ganz richtig, denn die Backpropagation ist mehr als die Anwendung des Gradientenabstiegs.

Bevor wir die Backpropagation erläutern, nochmal kurz zurück zur Forward-Propagation, die die eigentliche Prädiktion über ein künstliches neuronales Netz darstellt:

Forward-Propagation

Abbildung 1: Ein simples kleines künstliches neuronales Netz mit zwei Schichten (+ Eingabeschicht) und zwei Neuronen pro Schicht.

In einem kleinen künstlichen neuronalen Netz, wie es in der Abbildung 1 dargestellt ist, und das alle Neuronen über die Sigmoid-Funktion aktiviert, wird jedes Neuron eine Nettoeingabe z berechnen…

z = w^{T} \cdot x

… und diese Nettoeingabe in die Sigmoid-Funktion einspeisen…

\phi(z) = sigmoid(z) = \frac{1}{1 + e^{-z}}

… die dann das einzelne Neuron aktiviert. Die Aktivierung erfolgt also in der mittleren Schicht (N-Schicht) wie folgt:

N_{j} = \frac{1}{1 + e^{- \sum (w_{ij} \cdot x_{i}) }}

Die beiden Aktivierungsausgaben N werden dann als Berechnungsgrundlage für die Ausgaben der Ausgabeschicht o verwendet. Auch die Ausgabe-Neuronen berechnen ihre jeweilige Nettoeingabe z und aktivieren über Sigmoid(z).

Ausgabe eines Ausgabeknotens als Funktion der Eingänge und der Verknüpfungsgewichte für ein dreischichtiges neuronales Netz, mit nur zwei Knoten je Schicht, kann also wie folgt zusammen gefasst werden:

O_{k} = \frac{1}{1 + e^{- \sum (w_{jk} \cdot \frac{1}{1 + e^{- \sum (w_{ij} \cdot x_{i}) }}) }}

Abbildung 2: Forward-Propagation. Aktivierung via Sigmoid-Funktion.

Sollte dies die erste Forward-Propagation gewesen sein, wird der Output noch nicht auf den Input abgestimmt sein. Diese Abstimmung erfolgt in Form der Gewichtsanpassung im Training des neuronalen Netzes, über die zuvor erwähnte Gradientenmethode. Die Gradientenmethode ist jedoch von einem Fehler abhängig. Diesen Fehler zu bestimmen und durch das Netz zurück zu führen, das ist die Backpropagation.

Back-Propagation

Um die Gewichte entgegen des Fehlers anpassen zu können, benötigen wir einen möglichst exakten Fehler als Eingabe. Der Fehler berechnet sich an der Ausgabeschicht über eine Fehlerfunktion (Loss Function), beispielsweise über den MSE (Mean Squared Error) oder über die sogenannte Kreuzentropie (Cross Entropy). Lassen wir es in diesem Beispiel einfach bei einem simplen Vergleich zwischen dem realen Wert (Sollwert o_{real}) und der Prädiktion (Ausgabe o) bleiben:

e_{o} = o_{real} - o

Der Fehler e ist also einfach der Unterschied zwischen dem Ziel-Wert und der Prädiktion. Jedes Training ist eine Wiederholung von Prädiktion (Forward) und Gewichtsanpassung (Back). Im ersten Schritt werden üblicherweise die Gewichtungen zufällig gesetzt, jede Gewichtung unterschiedlich nach Zufallszahl. So ist die Wahrscheinlichkeit, gleich zu Beginn die “richtigen” Gewichtungen gefunden zu haben auch bei kleinen neuronalen Netzen verschwindend gering. Der Fehler wird also groß sein und kann über den Gradientenabstieg durch Gewichtsanpassung verkleinert werden.

In diesem Beispiel berechnen wir die Fehler e_{1} und e_{2} und passen danach die Gewichte w_{j,k} (w_{1,1} & w_{2,1} und w_{1,2} & w_{2,2}) der Schicht zwischen dem Hidden-Layer N und dem Output-Layer o an.

Abbildung 3: Anpassung der Gewichtungen basierend auf dem Fehler in der Ausgabe-Schicht.

Die Frage ist nun, wie die Gewichte zwischen dem Input-Layer X und dem Hidden-Layer N anzupassen sind. Es stellt sich die Frage, welchen Einfluss diese auf die Fehler in der Ausgabe-Schicht haben?

Um diese Gewichtungen anpassen zu können, benötigen wir den Fehler-Anteil der beiden Neuronen N_{1} und N_{2}. Dieser Anteil am Fehler der jeweiligen Neuronen ergibt sich direkt aus den Gewichtungen w_{j,k} zum Output-Layer:

e_{N_{1}} = e_{o1} \cdot \frac{w_{1,1}}{w_{1,1} + w_{1,2}} + e_{o2} \cdot \frac{w_{1,2}}{w_{1,1} + w_{1,2}}

e_{N_{2}} = e_{o1} \cdot \frac{w_{2,1}}{w_{2,1} + w_{2,2}} + e_{o2} \cdot \frac{w_{2,2}}{w_{2,1} + w_{2,2}}

Wenn man das nun generalisiert:

    \[ e_{N} = \left(\begin{array}{rr} \frac{w_{1,1}}{w_{1,1} + w_{1,2}} & \frac{w_{1,2}}{w_{1,1} + w_{1,2}} \\ \frac{w_{2,1}}{w_{2,1} + w_{2,2}} & \frac{w_{2,2}}{w_{2,1} + w_{2,2}} \end{array}\right) \cdot \left(\begin{array}{c} e_{1} \\ e_{2} \end{array}\right) \qquad \]

Dabei ist es recht aufwändig, die Gewichtungen stets ins Verhältnis zu setzen. Diese Berechnung können wir verkürzen, indem ganz einfach direkt nur die Gewichtungen ohne Relativierung zur Kalkulation des Fehleranteils benutzt werden. Die Relationen bleiben dabei erhalten!

    \[ e_{N} = \left(\begin{array}{rr} w_{1,1} & w_{1,2} \\ w_{2,1} & w_{2,2} \end{array}\right) \cdot \left(\begin{array}{c} e_{1} \\ e_{2} \end{array}\right) \qquad \]

Oder folglich in Kurzform: e_{N} = w^{T} \cdot e_{o}

Abbildung 4: Vollständige Gewichtsanpassung auf Basis der Fehler in der Ausgabeschicht und der Fehleranteile in der verborgenden Schicht.

Und nun können, basierend auf den Fehleranteilen der verborgenden Schicht N, die Gewichtungen w_{i,j} zwischen der Eingabe-Schicht I und der verborgenden Schicht N angepasst werden, entgegen dieser Fehler e_{N}.

Die Backpropagation besteht demnach aus zwei Schritten:

  1. Fehler-Berechnung durch Abgleich der Soll-Werte mit den Prädiktionen in der Ausgabeschicht und durch Fehler-Rückführung zu den Neuronen der verborgenden Schichten (Hidden-Layer)
  2. Anpassung der Gewichte entgegen des Gradientenanstiegs der Fehlerfunktion (Loss Function)

Buchempfehlungen

Die folgenden zwei Bücher haben mir sehr beim Verständnis und beim Verständlichmachen der Backpropagation in künstlichen neuronalen Netzen geholfen.

Neuronale Netze selbst programmieren: Ein verständlicher Einstieg mit Python Deep Learning. Das umfassende Handbuch: Grundlagen, aktuelle Verfahren und Algorithmen, neue Forschungsansätze (mitp Professional)

Training eines Neurons mit dem Gradientenverfahren

Dies ist Artikel 3 von 6 der Artikelserie –Einstieg in Deep Learning.

Das Training von neuronalen Netzen erfolgt nach der Forward-Propagation über zwei Schritte:

  1. Fehler-Rückführung über aller aktiver Neuronen aller Netz-Schichten, so dass jedes Neuron “seinen” Einfluss auf den Ausgabefehler kennt.
  2. Anpassung der Gewichte entgegen den Gradienten der Fehlerfunktion

Beide Schritte werden in der Regel zusammen als Backpropagation bezeichnet. Machen wir erstmal einen Schritt vor und betrachten wir, wie ein Neuron seine Gewichtsverbindungen zu seinen Vorgängern anpasst.

Gradientenabstiegsverfahren

Der Gradientenabstieg ist ein generalisierbarer Algorithmus zur Optimierung, der in vielen Verfahren des maschinellen Lernens zur Anwendung kommt, jedoch ganz besonders als sogenannte Backpropagation im Deep Learning den Erfolg der künstlichen neuronalen Netze erst möglich machen konnte.

Der Gradientenabstieg lässt sich vom Prinzip her leicht erklären: Angenommen, man stünde im Gebirge im dichten Nebel. Das Tal, und somit der Weg nach Hause, ist vom Nebel verdeckt. Wohin laufen wir? Wir können das Ziel zwar nicht sehen, tasten uns jedoch so heran, dass unser Gehirn den Gradienten (den Unterschied der Höhen beider Füße) berechnet, somit die Steigung des Bodens kennt und sich entgegen dieser Steigung unser Weg fortsetzt.

Konkret funktioniert der Gradientenabstieg so: Wir starten bei einem zufälligen Theta \theta (Random Initialization). Wir berechnen die Ausgabe (Forwardpropogation) und vergleichen sie über eine Verlustfunktion (z. B. über die Funktion Mean Squared Error) mit dem tatsächlich korrekten Wert. Auf Grund der zufälligen Initialisierung haben wir eine nahe zu garantierte Falschheit der Ergebnisse und somit einen Verlust. Für die Verlustfunktion berechnen wir den Gradienten für gegebene Eingabewerte. Voraussetzung dafür ist, dass die Funktion ableitbar ist. Wir bewegen uns entgegen des Gradienten in Richtung Minimum der Verlustfunktion. Ist dieses Minimum (fast) gefunden, spricht man auch davon, dass der Lernalgorithmus konvergiert.

Das Gradientenabstiegsverfahren ist eine Möglichkeit der Gradientenverfahren, denn wollten wir maximieren, würden wir uns entlang des Gradienten bewegen, was in anderen Anwendungen sinnvoll ist.

Ob als “Cost Function” oder als “Loss Function” bezeichnet, in jedem Fall ist es eine “Error Function”, aber auf die Benennung kommen wir später zu sprechen. Jedenfalls versuchen wir die Fehlerrate zu senken! Leider sind diese Funktionen in der Praxis selten so einfach konvex (zwei Berge mit einem Tal dazwischen).

 

Aber Achtung: Denn befinden wir uns nur zwischen zwei Bergen, finden wir das Tal mit Sicherheit über den Gradienten. Befinden wir uns jedoch in einem richtigen Gebirge mit vielen Bergen und Tälern, gilt es, das richtige Tal zu finden. Bei der Optimierung der Gewichtungen von künstlichen neuronalen Netzen wollen wir die besten Gewichtungen finden, die uns zu den geringsten Ausgaben der Verlustfunktion führen. Wir suchen also das globale Minimum unter den vielen (lokalen) Minima.

Programmier-Beispiel in Python

Nachfolgend ein Beispiel des Gradientenverfahrens zur Berechnung einer Regression. Wir importieren numpy und matplotlib.pyplot und erzeugen uns künstliche Datenpunkte:

Nun wollen wir einen Lernalgorithmus über das Gradientenverfahren erstellen. Im Grunde haben wir hier es bereits mit einem linear aktivierten Neuron zutun:

Bei der linearen Regression, die wir durchführen wollen, nehmen wir zwei-dimensionale Daten (wobei wir die Regression prinzipiell auch mit x-Dimensionen durchführen können, dann hätte unser Neuron weitere Eingänge). Wir empfangen einen Bias (w_0) der stets mit einer Eingangskonstante multipliziert und somit als Wert erhalten bleibt. Der Bias ist das Alpha \alpha in einer Schulmathe-tauglichen Formel wie y = \beta \cdot x + \alpha.

Beta \beta ist die Steigung, der Gradient, der Funktion.

Sowohl \alpha als auch \beta sind uns unbekannt, versuchen wir jedoch über die Betrachtung unserer Prädiktion durch Berechnung der Formel \^y = \beta \cdot x + \alpha und den darauffolgenden Abgleich mit dem tatsächlichen y herauszufinden. Anfangs behaupten wir beispielsweise einfach, sowohl \beta als auch \alpha seien 0.00. Folglich wird \^y = \beta \cdot x + \alpha ebenfalls gleich 0.00 sein und die Fehlerfunktion (Loss Function) wird maximal sein. Dies war der erste Durchlauf des Trainings, die sogenannte erste Epoche!

Die Epochen (Durchläufe) und dazugehörige Fehlergrößen. Wenn die Fehler sinken und mit weiteren Epochen nicht mehr wesentlich besser werden, heißt es, das der Lernalogorithmus konvergiert.

Als Fehlerfunktion verwenden wir bei der Regression die MSE-Funktion (Mean Squared Error):

MSE = \sum(\^y_i - y_i)^2

Um diese Funktion wird sich nun alles drehen, denn diese beschreibt den Fehler und gibt uns auch die Auskunft darüber, ob wie stark und in welche Richtung sie ansteigt, so dass wir uns entgegen der Steigung bewegen können. Wer die Regeln der Ableitung im Kopf hat, weiß, dass die Ableitung der Formel leichter wird, wenn wir sie vorher auf halbe Werte runterskalieren. Da die Proportionen dabei erhalten bleiben und uns quadrierte Fehlerwerte unserem menschlichen Verstand sowieso nicht so viel sagen (unser Gehirn denkt nunmal nicht exponential), stört das nicht:

MSE = \frac{\frac{1}{2} \cdot \sum(\^y_i - y_i)^2}{n}

MSE = \frac{\frac{1}{2} \cdot \sum(w^T \cdot x_i - y_i)^2}{n}

Wenn die Mathematik der partiellen Ableitung (Ableitung einer Funktion nach jedem Gradienten) abhanden gekommen ist, bitte nochmal folgende Regeln nachschlagen, um die nachfolgende Ableitung verstehen zu können:

  • Allgemeine partielle Ableitung
  • Kettenregel

Ableitung der MSD-Funktion nach dem einen Gewicht w bzw. partiell nach jedem vorhandenen w_j:

\frac{\partial}{\partial w_j}MSE = \frac{\partial}{\partial w} \frac{1}{2} \cdot \sum(\^y - y_i)^2

\frac{\partial}{\partial w_j}MSE = \frac{\partial}{\partial w} \frac{1}{2} \cdot \sum(w^T \cdot x_i - y_i)^2

\frac{\partial}{\partial w_j}MSE = \frac{2}{n} \cdot \sum(w^T \cdot x_i - y_i) \cdot x_{ij}

Woher wir das x_{ij} am Ende her haben? Das ergibt sie aus der Kettenregel: Die äußere Funktion wurde abgeleitet, so wurde aus \frac{1}{2} \cdot \sum(w^T \cdot x_i - y_i)^2 dann \frac{2}{n} \cdot \sum(w^T \cdot x_i - y_i). Jedoch muss im Sinne eben dieser Kettenregel auch die innere Funktion abgeleitet werden. Da wir nach w_j ableiten, bleibt nur x_ij erhalten.

Damit können wir arbeiten! So kompliziert ist die Formel nun auch wieder nicht: \frac{2}{n} \cdot \sum(w^T \cdot x_i - y_i) \cdot x_{ij}

Mit dieser Formel können wir unsere Gewichte an den Fehler anpassen: (f\nabla ist der Gradient der Funktion!)

w_j = w_j - \nabla MSE(w_j)

Initialisieren der Gewichtungen

Die Gewichtungen \alpha und \beta müssen anfänglich mit Werten initialisiert werden. In der Regression bietet es sich an, die Gewichte anfänglich mit 0.00 zu initialisieren.

Bei vielen neuronalen Netzen, mit nicht-linearen Aktivierungsfunktionen, ist das jedoch eher ungünstig und zufällige Werte sind initial besser. Gut erprobt sind normal-verteilte Zufallswerte.

Lernrate

Nur eine Kleinigkeit haben wir bisher vergessen: Wir brauchen einen Faktor, mit dem wir anpassen. Hier wäre der Faktor 1. Das ist in der Regel viel zu groß. Dieser Faktor wird geläufig als Lernrate (Learning Rate) \eta (eta) bezeichnet:

w_j = w_j - \eta \cdot \nabla MSE(w_j)

Die Lernrate \eta ist ein Knackpunkt und der erste Parameter des Lernalgorithmus, den es anzupassen gilt, wenn das Training nicht konvergiert.

Die Lernrate \eta darf nicht zu groß klein gewählt werden, da das Training sonst zu viele Epochen benötigt. Ungeduldige erhöhen die Lernrate möglicherweise aber so sehr, dass der Lernalgorithmus im Minimum der Fehlerfunktion vorbeiläuft und diesen stets überspringt. Hier würde der Algorithmus also sozusagen konvergieren, weil nicht mehr besser werden, aber das resultierende Modell wäre weit vom Optimum entfernt.

Beginnen wir mit der Implementierung als Python-Klasse:

Die Klasse sollte so funktionieren, bevor wir sie verwenden, sollten wir die Input-Werte standardisieren:

Bei diesem Beispiel mit künstlich erzeugten Werten ist das Standardisieren bzw. das Fehlen des Standardisierens zwar nicht kritisch, aber man sollte es sich zur Gewohnheit machen. Testweise es einfach mal weglassen 🙂

Kommen wir nun zum Einsatz der Klasse, die die Regression via Gradientenabstieg absolvieren soll:

Was tut diese Instanz der Klasse LinearRegressionGD nun eigentlich?

Bildlich gesprochen, legt sie eine Gerade auf den Boden des Koordinatensystems, denn die Gewichtungen werden mit 0.00 initialisiert, y ist also gleich 0.00, egal welche Werte in x enthalten sind. Der Fehler ist dann aber sehr groß (sollte maximal sein, im Vergleich zu zukünftigen Epochen). Die Gewichte werden also angepasst, die Gerade somit besser in die Punktwolke platziert. Mit jeder Epoche wird die Gerade erneut in die Punktwolke gelegt, der Gesamtfehler (über alle x, da wir es hier mit dem Batch-Verfahren zutun haben) berechnet, die Werte angepasst… bis die vorgegebene Zahl an Epochen abgelaufen ist.

Schauen wir uns das Ergebnis des Trainings an:

Die Linie sieht passend aus, oder? Da wir hier nicht zu sehr in die Theorie der Regressionsanalyse abdriften möchten, lassen wir das testen und prüfen der Akkuratesse mal aus, hier möchte ich auf meinen Artikel Regressionsanalyse in Python mit Scikit-Learn verweisen.

Prüfen sollten wir hingegen mal, wie schnell der Lernalgorithmus mit der vorgegebenen Lernrate eta konvergiert:

Hier die Verlaufskurve der Cost Function:

Die Kurve zeigt uns, dass spätestens nach 40 Epochen kaum noch Verbesserung (im Sinne der Gesamtfehler-Minimierung) erreicht wird.

Wichtige Hinweise

Natürlich war das nun nur ein erster kleiner Einstieg und wer es verstanden hat, hat viel gewonnen. Denn erst dann kann man sich vorstellen, wie ein einzelnen Neuron eines künstlichen neuronalen Netzes grundsätzlich trainiert werden kann.

Folgendes sollte noch beachtet werden:

  • Lernrate \eta:
    Die Lernrate ist ein wichtiger Parameter. Wer das Programmier-Beispiel bei sich zum Laufen gebracht hat, einfach mal die Lernrate auf Werte zwischen 10.00 und 0.00000001 setzen, schauen was passiert 🙂
  • Globale Minima vs lokale Minima:
    Diese lineare zwei-dimensionale Regression ist ziemlich einfach. Neuronale Netze sind hingegen komplexer und haben nicht einfach nur eine simple konvexe Fehlerfunktion. Hier gibt es mehrere Hügel und Täler in der Fehlerfunktion und die Gefahr ist groß, in einem lokalen, nicht aber in einem globalen Minimum zu landen.
  • Stochastisches Gradientenverfahren:
    Wir haben hier das sogenannte Batch-Verfahren verwendet. Dieses ist grundsätzlich besser als die stochastische Methode. Denn beim Batch verwenden wir den gesamten Stapel an x-Werten für die Fehlerbestimmung. Allerdings ist dies bei großen Daten zu rechen- und speicherintensiv. Dann werden kleinere Unter-Stapel (Sub-Batches) zufällig aus den x-Werten ausgewählt, der Fehler daraus bestimmt (was nicht ganz so akkurat ist, wie als würden wir den Fehler über alle x berechnen) und der Gradient bestimmt. Dies ist schon Rechen- und Speicherkapazität, erfordert aber meistens mehr Epochen.

Buchempfehlung

Die folgenden zwei Bücher haben mir bei der Erstellung dieses Beispiels geholfen und kann ich als hilfreiche und deutlich weiterführende Lektüre empfehlen:

 

Machine Learning mit Python und Scikit-Learn und TensorFlow: Das umfassende Praxis-Handbuch für Data Science, Predictive Analytics und Deep Learning (mitp Professional) Hands-On Machine Learning with Scikit-Learn and TensorFlow: Concepts, Tools, and Techniques for Building Intelligent Systems

 

Predictive maintenance in Semiconductor Industry: Part 1

The process in the semiconductor industry is highly complicated and is normally under consistent observation via the monitoring of the signals coming from several sensors. Thus, it is important for the organization to detect the fault in the sensor as quickly as possible. There are existing traditional statistical based techniques however modern semiconductor industries have the ability to produce more data which is beyond the capability of the traditional process.

For this article, we will be using SECOM dataset which is available here.  A lot of work has already done on this dataset by different authors and there are also some articles available online. In this article, we will focus on problem definition, data understanding, and data cleaning.

This article is only the first of three parts, in this article we will discuss the business problem in hand and clean the dataset. In second part we will do feature engineering and in the last article we will build some models and evaluate them.

Problem definition

This data which is collected by these sensors not only contains relevant information but also a lot of noise. The dataset contains readings from 590. Among the 1567 examples, there are only 104 fail cases which means that out target variable is imbalanced. We will look at the distribution of the dataset when we look at the python code.

NOTE: For a detailed description regarding this cases study I highly recommend to read the following research papers:

  •  Kerdprasop, K., & Kerdprasop, N. A Data Mining Approach to Automate Fault Detection Model Development in the Semiconductor Manufacturing Process.
  • Munirathinam, S., & Ramadoss, B. Predictive Models for Equipment Fault Detection in the Semiconductor Manufacturing Process.

Data Understanding and Preparation

Let’s start exploring the dataset now. The first step as always is to import the required libraries.

There are several ways to import the dataset, you can always download and then import from your working directory. However, I will directly import using the link. There are two datasets: one contains the readings from the sensors and the other one contains our target variable and a timestamp.

The first step before doing the analysis would be to merge the dataset and we will us pandas library to merge the datasets in just one line of code.

Now let’s check out the distribution of the target variable

Figure 1: Distribution of Target Variable

From Figure 1 it can be observed that the target variable is imbalanced and it is highly recommended to deal with this problem before the model building phase to avoid bias model. Xgboost is one of the models which can deal with imbalance classes but one needs to spend a lot of time to tune the hyper-parameters to achieve the best from the model.

The dataset in hand contains a lot of null values and the next step would be to analyse these null values and remove the columns having null values more than a certain percentage. This percentage is calculated based on 95th quantile of null values.

Figure 2: Missing percentge in each column

Now we calculate the 95th percentile of the null values.

Figure 3: Missing percentage after removing columns with more then 45% Na

From figure 3 its visible that there are still missing values in the dataset and can be dealt by using many imputation methods. The most common method is to impute these values by mean, median or mode. There also exist few sophisticated techniques like K-nearest neighbour and interpolation.  We will be applying interpolation technique to our dataset. 

To prepare our dataset for analysis we should remove some more unwanted columns like columns with near zero variance. For this we can calulate number of unique values in each column and if there is only one unique value we can delete the column as it holds no information.

We have applied few data cleaning techniques and reduced the features from 590 to 444. However, In the next article we will apply some feature engineering techniques and adress problems like the curse of dimensionality and will also try to balance the target variable.

Bleiben Sie dran!!

The Inside Out of ML Based Prescriptive Analytics

With the constantly growing number of data, more and more companies are shifting towards analytic solutions. Analytic solutions help in extracting the meaning from the huge amount of data available. Thus, improving decision making.

Decision making is an important aspect of businesses, and technologies like Machine Learning are enhancing it further. The growing use of Machine Learning has changed the way of prescriptive analytics. In order to optimize the efforts, companies need to be more accurate with the historical and present data. This is because the historical and present data are the essentials of analytics. This article helps describe the inside out of Machine Learning-based prescriptive analytics.

Phases of business analytics

Descriptive analytics, predictive analytics, and prescriptive analytics are the three phases of business analytics. Descriptive analytics, being the first one, deals with past performance. Historical data is mined to understand past performance. This serves as a way to look for the reasons behind past success and failure. It is a kind of post-mortem analysis and most management reporting like sales, marketing, operations, and finance etc. make use of this.

The second one is a predictive analysis which answers the question of what is likely to happen. The historical data is now combined with rules, algorithms etc. to determine the possible future outcome or likelihood of a situation occurring.

The final phase, well known to everyone, is prescriptive analytics. It can continually take in new data and re-predict and re-prescribe. This improves the accuracy of the prediction and prescribes better decision options.  Professional services or technology or their combination can be chosen to perform all the three analytics.

More about prescriptive analytics

The analysis of business activities goes through many phases. Prescriptive analytics is one such. It is known to be the third phase of business analytics and comes after descriptive and predictive analytics. It entails the application of mathematical and computational sciences. It makes use of the results obtained from descriptive and predictive analysis to suggest decision options. It goes beyond predicting future outcomes and suggests actions to benefit from the predictions. It shows the implications of each decision option. It anticipates on what will happen when it will happen as well as why it will happen.

ML-based prescriptive analytics

Being just before the prescriptive analytics, predictive analytics is often confused with it. What actually happens is predictive analysis leads to prescriptive analysis. Thus, a Machine Learning based prescriptive analytics goes through an ML-based predictive analysis first. Therefore, it becomes necessary to consider the ML-based predictive analysis first.

ML-based predictive analytics:

A lot of things prevent businesses from achieving predictive analysis capabilities.  Machine Learning can be a great help in boosting Predictive analytics. Use of Machine Learning and Artificial Intelligence algorithms helps businesses in optimizing and uncovering the new statistical patterns. These statistical patterns form the backbone of predictive analysis. E-commerce, marketing, customer service, medical diagnosis etc. are some of the prospective use cases for Machine Learning based predictive analytics.

In E-commerce, machine learning can help in predicting the usual choices of the customer. Thus, presenting him/her according to his/her likes and dislikes. It can also help in predicting fraudulent transaction. Similarly, B2B marketing also makes good use of Machine learning based predictive analytics. Customer services and medical diagnosis also benefit from predictive analytics. Thus, a prediction and a prescription based on machine learning can boost various business functions.

Organizations and software development companies are making more and more use of machine learning based predictive analytics. The advancements like neural networks and deep learning algorithms are able to uncover hidden information. This all requires a well-researched approach. Big data and progressive IT systems also act as important factors in this.