Contextual Multi-Arm Bandit and its applications to digital experiments
summary
Claims a production-ready contextual MAB that builds a model from historical randomized data then assigns users to contextual cohorting and runs a vanilla MAB per cohort, implemented in a developing Python package optumwift. Methods include Thompson sampling (Bayesian logistic weights) and linear UCB baselines. Results come from simulation and held‑out A/B test traces showing contextual cohorts separate by t≈5–6 and outperform vanilla MAB and A/B testing; one cluster’s spike reflects tiny sample size.
AI-written summary · about
chapters (9)
- 0:00Contextual multi-arm bandits intro
- 7:26Vanilla bandit methods overview
- 18:45Why contextual bandits
- 22:11Existing methodologies
- 26:02Our implementation
- 27:26Proof of concept results
- 35:56Broader applications
- 37:40OptumWift Python package
- 39:03Q&A
related talks
41mKeynote: Personalization, Bandits, and Causal inference
48mAdaptivity and Confounding in Multi-armed Bandit Experiments
21mLeveraging Multi-Armed Bandit Algorithms for Dynamic Decision Making
20mWinning the Game: Leveraging Contextual Bandits for Creative Optimization
22mPersonalizing Explainable Recommendations with Multi-objective Contextual Bandits
32mMulti-Armed Bandits in Recommendations
33mTwo Years of Bayesian Bandits for E Commerce
01h 01mStochastic Bandits: Foundations and Current Perspectives