John Maxwell, Data Scientist, Nordstrom at MLconf Seattle 2017
summary
Contextual bandits replace slow A/B tests by routing live traffic with epsilon-greedy exploration while learning offline via a reduction to cost‑sensitive classification. Implementation uses a Dora node app that logs context, chosen arm and play probability, serves recommendations from TensorFlow Serving, and retrains in batches using inverse propensity weighting to debias the logged data. This pipeline avoids per-request matrix inversions, runs in production at Nordstrom, and converges far faster than standard A/B testing.
AI-written summary · about
chapters (7)
- 0:00Intro
- 0:40Product Recommendations
- 3:09Exploration Policy
- 4:12Confidence Bound
- 7:53Delayed Rewards
- 10:53Offline Classification
- 12:50Questions
related talks
17mByron Galbraith, Chief Data Scientist, Talla, NYC 2017
41mKeynote: Personalization, Bandits, and Causal inference
39mGrowing user engagement with RL-driven personalisation
20mWinning the Game: Leveraging Contextual Bandits for Creative Optimization
21mBuild, Measure, and Machine Learn with Angular
33mTwo Years of Bayesian Bandits for E Commerce
42mA/B testing with real-world data to evaluate ML model updates
22mPersonalizing Explainable Recommendations with Multi-objective Contextual Bandits