Keynote: Personalization, Bandits, and Causal inference
summary
Sequential experimentation with bandit methods beats naive A/B testing: Thompson sampling and UCB rapidly reduce regret versus epsilon‑first. Contextual models plus scheduled exploration let policies learn per‑person actions, while offline evaluation using known or estimated propensity scores and doubly robust estimators recovers randomized results (flexible additive‑tree models are used). Real examples include oncology registries (confounding makes naive models recommend withholding chemotherapy) and a clinical dataset where personalized assignment lowers adverse outcome rates by about 2%. An extensible contextual‑bandit server and an R package support these pipelines.
AI-written summary · about
chapters (11)
- 0:00Personalization research intro
- 3:20Defining personalization problem
- 9:15Bandits for personalization
- 14:11Exploration vs exploitation
- 21:43Contextual bandits extension
- 24:53Personalization from observational data
- 27:00Supervised learning in healthcare
- 30:12Offline policy evaluation
- 33:21Policy evaluation without propensities
- 36:13Causal discovery for personalization
- 37:17Closing remarks
related talks
45mContextual Multi-Arm Bandit and its applications to digital experiments
48mAdaptivity and Confounding in Multi-armed Bandit Experiments
20mWinning the Game: Leveraging Contextual Bandits for Creative Optimization
21mLeveraging Multi-Armed Bandit Algorithms for Dynamic Decision Making
22mPersonalizing Explainable Recommendations with Multi-objective Contextual Bandits
17mByron Galbraith, Chief Data Scientist, Talla, NYC 2017
32mDesigning Adaptive Experiments For Policy Learning And Inference
39mGrowing user engagement with RL-driven personalisation