Discovering Customer Journey Maps using a Mixture of Markov Models

Transcription

1 Discovering Customer Journey Maps using a Mixture of Markov Models SIMPDA Neuchâtel (Switzerland) Matthieu Harbich - Presenter Gaël Bernard - University of Lausanne Dr. Pietro Berkes - Kudelski Pr. Benoît Garbinato Pr. Periklis Andritsos

2 map Customer Journeys. Path followed by a customer to consume a service. CAR DEALER 1

3 map Customer Journeys: Capture the Customer Experience. s are becoming increasingly complex Numerous channels / devices Increasing number of interactions «Improving understanding of customers has been ranked one of the most important research challenges in the coming year [1]» [1] Research priorities Tech. rep., Marketing Science Institute (2016), 2

4 CJMs: A customer-centric approach. map Customer Journey Maps (CJMs) are used to understand, discuss, or improve the main paths in the usage of a service Mostly used as a design thinking tool / strategic tool Getting information Trying Buying Sharing experience 3

5 Strategy vs. Reality. map «People don t behave like robots, and no matter how well we craft an experience, they will not perceive exactly as we anticipate or hope» [2] [2] Richardson, A.: Series on customer journey: Using customer journey maps to improve customer experience, using customer journey maps to improve customer experience, touchpoint bring the customer experience to life. Harb Bus Rev (2010) 4

6 XES: Discovering CJMs from Event Logs. map When customers interact with a service, it (often) leaves traces We can leverage these traces to provide fact-based insights on customer journeys Inspired from Process Mining INPUT: The XES standard is adequate to store CJMs [3] [3] Bernard, G., & Andritsos, P. (2017). A Process Mining Based Model for Customer Journey Mapping. 5

7 : Summarize the Logs using K Representative Journeys. map Event logs (e.g., XES) B D A E C G A B D A F B D F E C G C A A B C D E F G CJM with actual journeys time A B C D E F G CJM with representative journeys 1, 3, 4 5, 2 time 6

8 : Summarize the Logs using K Representative Journeys. map Event logs (e.g., XES) B D A E C G A B D A F B D F E C G C A A B C D E F G CJM with actual journeys time A B C D E F G CJM with representative journeys 1, 3, 4 5, 2 time 6

9 : Summarize the Logs using K Representative Journeys. map Event logs (e.g., XES) B D A E C G A B D A F B D F E C G C A A B C D E F G CJM with actual journeys time A B C D E F G CJM with representative journeys 1, 3, 4 5, 2 time 6

10 Markov models: Modelling sequences. map «In probability theory, a Markov model is a stochastic model used to model randomly changing systems. It is assumed that future states depend only on the current state, not on the events that occurred before it.» - Paul A. Gagniuc a b c d Markov model c d b c d a a b c d a b c d First probability vector Transition matrix Gagniuc, Paul A. (2017). Markov Chains: From Theory to Implementation and Experimentation. USA, NJ: John Wiley & Sons. pp ISBN

11 Expectation-Maximization on a Mixture of Markov Models. map Input Algorithm Output Expectation XES Learning the parameters of the K Markov models K Markov models K Maximization Optimizing the log-likelihood of the data 8

12 : The Dataset. map An activity-based travel survey conducted in the Chicago metropolitan area over a demographic representative sample of its population. 29,542 journeys - 2,381 are unique being at home going to school having a meal being at home 16 types of events Average number of events per journey:

13 Case Study: Results. map 10

14 : Each Model is Responsible for each Journey. map 11

15 : Each Model is Responsible for each Journey. map 11

16 : Each Model is Responsible for each Journey. map 11

17 : Each Model is Responsible for each Journey. map 12

18 : Each Model is Responsible for each Journey. map 12

19 map : Limited Number of Parameters. As input: X Feature engineering X Distance metric K - Which can potentially be computed upfront thanks to information criterion technics Parameters set by a human can have an important impact on the representative journeys. The E-M on a mixture of Markov models allows to discover the representative journeys much more naturally. 13

20 : Computing the Next Most Probable Event. map a b c d a b c d 14

21 : Handling Shifting Behaviors - Soft Clustering. map c T0 Model 1 T0 Model 3 Model 2 15

22 : Handling Shifting Behaviors - Soft Clustering. map c T0 d T1 Model 1 T1 Model 3 Model 2 15

23 : Handling Shifting Behaviors - Soft Clustering. map c d b T0 T1 T2 Model 1 T2 Model 3 Model 2 15

24 : Handling Shifting Behaviors - Soft Clustering. map c d b c T0 T1 T2 T3 Model 1 T3 Model 3 Model 2 15

25 map Kudelski application Probabilistic approach robustness Research in progress! Q&A