Estimating the Value of Personalization
Talkat Bridging Prediction and Intervention Problems in Social Systems (Simons Institute, 2026) · 2026-01-12..01-18·
summary
Introduces a new estimator that, from historical data, tests whether a personalized policy yields higher expected outcomes than the best one-size-fits-all action and provides a p-value; uses a repeated-splits k-fold personalization test with doubly robust AIPW estimator to avoid optimistic bias from reusing data, proves zero asymptotic Type I error and efficiency under fast best-arm and consistent policy-learning rates, and shows on JobCorps and a depression RCT that the estimator is more stable and less split-sensitive than train/eval or Xie et al baselines.
AI-written summary · about
chapters (6)
- 0:00Personalization in AI systems
- 1:52Why personalization matters
- 7:46Mathematical personalization effect
- 15:43K-fold personalization test
- 25:48Empirical evaluation JobCorps data
- 31:23Future directions
related talks
41mKeynote: Personalization, Bandits, and Causal inference
30mTalk by Suhas Diggavi (UCLA)
41mConfident Off-policy Evaluation and Selection through Self-Normalized Importance Weighting
01h 02mEstimation Beyond the IID Setting (Two Vignettes)
47mPrivate targeting strategies
12mCausal Inference And Data-Driven Decision-Making
57mA New Perspective on High-Dimensional Causal Inference
09mTowards Credible and Effective Data-Driven Decision-Making