advertisement

advertisement

Information about RepèRes Bayesia Consumer Segmentation Skim Conf08

Published on May 30, 2008

Author: f.abiven

Source: slideshare.net

advertisement

Summary 2 Introduction to consumer segmentations A brief overview of Bayesian Networks Computing a segmentation with Bayesian Networks Conclusion Skim Conference – Barcelona – May 28th 2008

Introduction to consumer segmentations 3 Introduction to consumer segmentations A brief overview of Bayesian Networks Computing a segmentation with Bayesian Networks Conclusion Skim Conference – Barcelona – May 28th 2008

Why a segmentation ? 4 Valuable tool to understand a market Homogeneous marketing targets - people who behave the same way - people who have homogeneous motivations / attitudes. Groups of people to whom it is possible to speak the same language Different marketing strategies # Concepts # Products # Communication # Advertising MORE EFFICIENT Skim Conference – Barcelona – May 28th 2008

A good segmentation - some important features 5 Homogeneous segments TECHNICAL QUALITY Clear differences between segments Stable… AND OTHER Easy to understand VERY IMPORTANT Operational / Actionable ELEMENTS Fair representation of the real world Preparation Statistical Interpretation Output stage procedure / Analysis Only a part of the whole process. How important is it ? Skim Conference – Barcelona – May 28th 2008

The marketer’s dream…and cruel reality 6 Obvious groups ! More complicated Any kind of computation should Unlimited number of typologies lead to the same results Procedure should guarantee a relevant clustering Skim Conference – Barcelona – May 28th 2008

Classical procedures 7 A factorial analysis followed by a clustering of the individuals Canonical segmentation ATTITUDES ATTITUDES BEHAVIOURS BEHAVIOURS CANONICAL ANALYSIS CANONICAL ANALYSIS Projection of the individuals on the factorial axis Projection of the individuals on the factorial axis Clustering of the individuals Clustering of the individuals Drawbacks : Difficult to choose what are the attitudes / what are the behaviours (declarative statements) – Time consuming. Skim Conference – Barcelona – May 28th 2008

A brief overview of Bayesian Networks 8 Introduction to consumer segmentations A brief overview of Bayesian Networks Computing a segmentation with Bayesian Networks Conclusion Skim Conference – Barcelona – May 28th 2008

Bayesian Networks 9 A computational Tool to Model Uncertainty based both on graphs theory readability – Powerful communication tool and probability theory sound computations Manual modelling through brainstorming Probabilistic Expert Systems Induction by automatic learning Data analysis, data mining Growing popularity Industry, Defense, Health, …and now, Market Research Skim Conference – Barcelona – May 28th 2008

A complete framework for Data Mining 10 Parametric estimation Use of the database to estimate the probabilities of a given structure Robust Missing values processing Expectation-Maximization (EM) Structural EM Structural learning Unsupervised learning to discover all the direct probabilistic relations Supervised learning to characterize a target variable Variable clustering to induce “factors” made of highly connected variables Probabilistic Structural Equations and… Data Clustering to find groups of data sharing the same characteristics Skim Conference – Barcelona – May 28th 2008

Formalism : 2 distinctive parts 11 Structure Directed acyclic graphs Example: Anti-doping Parameters agency using two Probability distributions associated to each node different tests to screen competitors Skim Conference – Barcelona – May 28th 2008

A reasoning engine 1/3 12 Sound evidence propagation on the entire network Simulation Diagnosis And any combination of these 2 types of inference Skim Conference – Barcelona – May 28th 2008

A reasoning engine 2/3 13 Sound evidence propagation on the entire network Simulation Diagnosis If a competitor is doped... …there is 99.5% chance that he is disqualified Skim Conference – Barcelona – May 28th 2008

A reasoning engine 3/3 14 Sound evidence propagation on the entire network Simulation Diagnosis : thinking the other way round … there is a slight probability (8%) that he is nevertheless clean. If a competitor has been disqualified… Skim Conference – Barcelona – May 28th 2008

Segmentation with Bayesian Networks 15 Introduction to consumer segmentations A brief overview of Bayesian Networks Computing a segmentation with Bayesian Networks Real case study: Segmentation of women as regards shopping and fashion For confidentiality reasons, consumer statements and outputs have been modified. Conclusion Skim Conference – Barcelona – May 28th 2008

1st Stage : segmentation induction 16 Skim Conference – Barcelona – May 28th 2008

Unsupervised learning 17 Discovering relations between consumer statements Usage and attitude survey conducted for a clothes retailer. Sample=1065 women. 234 consumer statements: attitudes and behaviours towards fashion in general, retailers, brand image… Heuristic Search Algorithm to find the best representation of the joint probability distribution. Minimum Description Length score to evaluate the quality of the network based on fitness and compactness Induced network Skim Conference – Barcelona – May 28th 2008

Variables clustering and factor induction 18 Simplifying the information Analysis of the network to discover groups of variables that are strongly connected and that form a “concept” Ascendant Hierarchical Clustering algorithm based on the arcs’ Kullback Leibler forces (non linear and global measure – contribution of the relation to the network). For each cluster of variables Creation of a latent variable summarizing the information. 42 factors computed Example of factor 15 : dimension summarizing originality. Based on attitude statements Latent variable (importance to be original, like to differentiate with clothes) and behaviours (buy brands X, Y and Z more often). Skim Conference – Barcelona – May 28th 2008

Factor clustering: overview of the procedure 19 Segmentation of the individuals based on the main factors Introducing a new variable (consumer segments) which is the hidden cause of the main factors. Learning the probabilities with Expectation – Maximisation Score derived from MDL to assess the quality of the clustering Skim Conference – Barcelona – May 28th 2008

Selecting the number of clusters 20 Pseudo random walk to find the best number of clusters example: find the best clustering with random walk between 2 and 6 clusters – 20 iterations The best segmentation is the one that minimizes the score Also possible to define the desired number of clusters Possible to define the minimal purity of the clusters. The purity is computed as the mean of the probability of each cluster point. Skim Conference – Barcelona – May 28th 2008

2nd stage : segmentation analysis 21 Skim Conference – Barcelona – May 28th 2008

Supervised learning 22 Focusing on consumer clusters LEARNING the relations between… THE TARGET VARIABLE = SEGMENTATION THE CONSTITUTIVE VARIABLES = CONSUMER STATEMENTS Target Variable = consumer segments Skim Conference – Barcelona – May 28th 2008

Cluster Profile 23 Using the network to describe the consumer groups Identification of the key variables and associated values For each consumer group, we use the % of shared information to sort the variables according to their importance in the characterisation of the group. 4 most contributing variables Compared with total sample, women of cluster#5 : for Cluster #5 - Buy brand X more often - Are older women (59 in average) - Do not consider originality as important - Do not like discovering new shops Arrows symbolize the change in the probability distribution when observing cluster #5. Skim Conference – Barcelona – May 28th 2008

Generation of the cluster mapping 24 Map generation The size of the cluster is proportional to its probability The proximity of the clusters is a probabilistic proximity The darkness of the blue is proportional to the purity of the cluster (in this example all clusters have a purity > 95%) Skim Conference – Barcelona – May 28th 2008

Summarizing segmentation results 25 -- Money devoted to clothes 18% 10% Fashion cheap Functional before all above all 20% Age Fashionable Neutral Classical originality 18% Superstars 20% 8% 14% Classical upmarket Young manager / executive ++ Money women devoted to clothes Skim Conference – Barcelona – May 28th 2008

Going further : identifying a more compact target model 26 Markov procedure to select a subset of statements to determine to which category consumers belong Selection of a subset of variables… …knowing the values of these variables makes the target independent of all the other variables Subset of 11 variables Overall prediction score = 68% Interesting to quickly recruit consumer groups amongst the total population. Skim Conference – Barcelona – May 28th 2008

Conclusion 27 Introduction to consumer segmentations A brief overview of Bayesian Networks Computing a segmentation with Bayesian Networks Conclusion Skim Conference – Barcelona – May 28th 2008

Benefits 28 Our experience : a powerful tool - Relevant typologies - Easy to carry out Modelling the consumer variables : good representation of reality - Non-supervised modelling : no strong hypothesis - Discovering interactions between variables (behaviours / attitudes) - Use of qualitative / quantitative variables Data clustering quality - Possible to set the minimum purity of the clusters : enables the marketer to discover “niche” markets (usually less pure) or focus on mainstream groups. Added-value in the analysis of the clusters - Easy ranking of the key variables for each consumer cluster - Proximity mapping to summarize results Development of robust models to identify consumer groups - Interesting in the case of upcoming recruitment. Skim Conference – Barcelona – May 28th 2008

Some drawbacks. How to deal with them ? 29 Modelling the consumer network and computing latent variables can be long when the number of variables is very important. 234 variables and 1065 lines: 30-40 minutes To speed up the process, possible to learn a simplified network : e.g. maximum spanning tree or increase of the structural complexity parameter. Continuous variables have to be discretized Results will depend on the quality of the discretization. Possible to use K-Means to adapt discretization to the distribution of the data. Expertise of the user also helps. And most of the time in consumer research variables are discrete ! Skim Conference – Barcelona – May 28th 2008

Perspectives 30 Flexibility : can be used far beyond usage and attitudes surveys Easy to carry out Can be adapted to any type of data Well designed to process large amounts of data Example: segmentation of trains using client’s internal data Travelers' Data Train data (turnover, occupancy rate…) 10 Million individuals 15.000 trains Clustering of trains In the future… - typology of clients (turnover, potential…) to feed a business strategy - segmentation of consumers based on utilities (CBC data) Skim Conference – Barcelona – May 28th 2008

Contact 31 Jouffe Lionel Craignou Fabien Managing Director Data Mining Department Manager jouffe@bayesia.com fcr@reperes.net Skim Conference – Barcelona – May 28th 2008

Repères Articles, experts, jobs, and more: get all the professional insights you need on LinkedIn. Share on LinkedIn Share on Twitter Share on Facebook ...

Read more

Consumer Segmentation at a glance: 6,607 LinkedIn members have this skill. ... Associate Director of Consumer Insights at Burrell Communications Past

Read more

blog of the CEO of Reperes ... I’m a consumer: ... From this perspective which segmentation do you feel is most relevant? YB: ...

Read more

... Bayesia – Repères. Bayesian Networks: a new tool for consumer segmentation. In Skim Conference, Barcelona, May 28th 2008. [4] ...

Read more

## Add a comment