Two Years of Bayesian Bandits for E Commerce
summary
Monetate deploys Bayesian bandits since early 2016 to optimize e‑commerce; implementation uses a beta‑binomial likelihood and Thompson sampling with Monte Carlo approximations but scales by sampling once per decision so JS tag latency meets a 1–2 ms prediction budget. Simulations show fast allocation to the 10% vs 5% arm, unexpected behavior on AAA tests (wide, nonshrinking intervals), and systematic bias in observed rates; inverse‑propensity weighting debiases estimates at the cost of much higher variance, creating concrete UX and trust tradeoffs for marketers.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 3:54Web optimization basics
- 8:09Sequential testing to bandits
- 10:39Bandit implementation basics
- 16:55Thompson sampling trick
- 17:58Simulating bandit behavior
- 20:05AAA bandit surprises
- 22:56Bias in reward estimates
- 27:33UX challenges solutions
- 28:23Q&A
related talks
17mByron Galbraith, Chief Data Scientist, Talla, NYC 2017
21mLeveraging Multi-Armed Bandit Algorithms for Dynamic Decision Making
41mKeynote: Personalization, Bandits, and Causal inference
30mPractical Multi Armed Bandits
38mUsable A/B testing – A Bayesian approach
20mWinning the Game: Leveraging Contextual Bandits for Creative Optimization
32mMulti-Armed Bandits in Recommendations
45mContextual Multi-Arm Bandit and its applications to digital experiments