"Haute Couture" and "Prêt-à-Porter" Data Science

Talk given @ Telecom ParisTech on April 2016

1. Christophe Bourguignat zelros.com / human@zelros.com / @zelrosHQ
3. Agenda Models interpretation Models production A short history of Kaggle
4. MODELS INTERPRETATION
5. WHY ? Models opacity is a major reject cause by users Unfortunately, predictive models that are the most powerful are usually the least interpretable
9. FEATURE IMPORTANCE
13. AEROSOLVE (AirBnb) Prior = general belief, before looking at the data Inform the model of our prior beliefs by adding them to a text configuration file during training
17. Scikit Learn
18. Scikit Learn March 2014
19. Scikit Learn March 2014 April 2015
20. Scikit Learn March 2014 April 2015
21. Scikit Learn March 2014 April 2015
22. Scikit Learn March 2014 April 2015
23. Scikit Learn https://github.com/andosa/treeinterpreter/blob/master/treeinterpreter/treeinterpreter.py
24. EXEMPLE ON BOSTON DATASET
26. http://blog.datadive.net/prediction-intervals-for-random-forests/ Prediction Intervals for Random Forests
29. PRODUCTION
32. TRADITIONAL B.I. DEPARTMENT DATA ANALYSTS ETL ENGINEER DBAs
33. “INFINITE LOOP OF SADNESS” DATA SCIENTISTS IT / DATA ENGINEERS SOFTWARE ENGINEERS BUSINESS http://multithreaded.stitchfix.com/blog/2016/03/16/engineers-shouldnt-write-etl/
34. CODE http://treycausey.com/software_dev_skills.html
35. COMPLEXITY AND TECHNICAL DEBT Underutilized features Undeclared consumers Pipeline Jungles - preparing data in a ML-friendly format http://static.googleusercontent.com/media/research.google.com/fr//pubs/archive/43146.pdf
36. PRODUCTION FAILS Unseen category Unreproductible feat eng workflow (PMML) Leakage in DataBase fields (churn) Monitoring
37. A BRIEF HISTORY OF KAGGLE
38. June 2013 Sept 2013 Nov 2014 Apr 2015 Mar 2016
46. Refinements : - hashing function - adaptive learning rate (different flavours) - Vowpal Wabbit - Dropout - PyPy
55. QUESTIONS ? zelros.com / human@zelros.com / @zelrosHQ