Winning (Some) Churners Back
summary
Booking.com builds a churn-score pipeline and an experimentation framework that targets incentives to users by score rather than blasting everyone; early model is a random forest (AUC ~0.75), users are split into five-probability buckets for A/B tests, and untargeted incentives show heavy cannibalization. Some middle buckets deliver ROI > 1 while others lose money. Next step switches to a Cox proportional hazards model to score near-term likelihoods and tune who actually receives offers.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 2:12Incentive campaigns for churners
- 5:32Predictive modeling improves ROI
- 9:36Testing model predictions
- 14:58Continuous improvement framework
- 17:09Survival models for churn
- 24:19Addressing audience questions on cannibalization
- 26:29Evolving churn probabilities
- 28:43Optimizing incentive amounts
related talks
27mPreventing churn like a bandit
24mUnlocking revenue: newcomer classification for boosted profits
33mCausal Inference Framework for incrementality
24mOffline Ranking Validation - Predicting A/B Test Results
32mClassifying the variety of customers online engagement for churn prediction
45mKnow Your Customer
20mWinning the Game: Leveraging Contextual Bandits for Creative Optimization
20mUsers acquired - Now what?: Maximising revenue from existing players