Tree Tricks on Customer Segmentation

Lightning Talk at EuroSciPy @Cambridge

Transcript

1. 1 2015 © All rights reserved to SciPy Lighting Talk: Three Tricks on Customer Segmentation Volodymyr (Vlad) Kazantsev Head of Data Science at Product Madness jobs@productmadness.com volodymyrk
2. 2 volodymyrk Heart of Vegas in (public) Numbers iPad US - #13 top grossing iPhone US - #32 top grossing Android - #44 top grossing US (games) Australia iPad - #1 top grossing iPhone - #1 top grossing Android -#3 top grossing
3. 3 volodymyrk Data Impact Team • Ad-hoc analytics and daily fires; dashboards • Deep dive analysis; Predictive analytics • ETL, Data Viz tools, R&D, DBA Analytics Data Science Data Engineering 7 people; 4 in London office
4. 4 volodymyrk What is a Player Segment? A segment is a group of customers who display similar attributes to each other... Customers move in and out of segments over time
5. 5 volodymyrk How to actually do segmentation? Just Look at Data Clustering Decision Trees Player Attributes de-correlate Normalise Scale
6. 6 De-correlate features 1
7. 7 volodymyrk normalise and de-correlate Player 1 more similar to Player 2 ? Player 3 more similar to Player 2 ? Weekly Play Summary
8. 8 volodymyrk normalise and de-correlate Weekly Play Summary
9. 9 volodymyrk normalise and de-correlate Player 1 more similar to Player 2 !
10. 10 volodymyrk normalise and de-correlate Player 1 more similar to Player 2 !
11. 11 volodymyrk normalise and de-correlate Player 1 more similar to Player 2, isn’t he?
12. 12 volodymyrk normalise and de-correlate Player 3 more similar to Player 2 !
13. 13 Decision Trees .. for Segmentation? 2
14. 14 volodymyrk What now? • K-means • Hierarchical Clustering • Decision Trees • .. and many more
15. 15 volodymyrk Decision Tree for Clustering All Payers 500 (next month>$100): 4.7% 10000 did not: 95.3% Last_months_dollars <=$2 2 (next month>$100): 0.04% 5000 did not: 99% Last_months_dollars >$2 498 (next month>$100) > $100: 9% 5000 did not: 91% Transactions <=10 243 (next month>$100): 5.5% 4200 did not: 94.5% Transactions > 10 255 (next month>$100): 24% 800 did not: 76%
16. 16 volodymyrk Decision Tree for Clustering All Payers 500 (next month>$100): 4.7% 10000 did not: 95.3% Last_months_dollars <=$2 2 (next month>$100): 0.04% 5000 did not: 99% Last_months_dollars >$2 498 (next month>$100) > $100: 9% 5000 did not: 91% Transactions <=10 243 (next month>$100): 5.5% 4200 did not: 94.5% Transactions > 10 255 (next month>$100): 24% 800 did not: 76% Low Value Medium Value High Value
17. 17 Segmentation in A/B tests 3
18. 18 volodymyrk Segmentation for A/B tests A B
19. 19 volodymyrk Segmentation for A/B tests A B
20. 20 volodymyrk Bonferroni correction: Bayesian Hierarchical Model Combine stats with Market Intuition! Adjustment for multiple testing adjustted = desired /M
21. 21 2015 © All rights reserved to Thank You! jobs.productmadness.com volodymyr.kazantsev@productmadness.com volodymyrk We Are Hiring ! jobs.productmadness.com