Byron Galbraith, Chief Data Scientist, Talla, NYC 2017
summary
Builds a realtime recommender for a VR silverware shopping service and claims Bayesian bandits find better options faster than A/B tests. Uses Thompson sampling with a Beta–Bernoulli conjugate (tracking successes and failures per arm, initialized 2 and 1), models binary engagement, and shows price drops conversion from 50% to 22% while an association/design option rises to ~60–70%. Bandits converge within hundreds of trials, cut cumulative regret versus fixed splits, but need contextual handling for nonstationary traffic.
AI-written summary · about
chapters (8)
- 0:00Start
- 3:51Reinforcement Learning
- 6:46Multi-Armed Bandit Problem
- 7:47Michael Thompson Sampling
- 9:01The Bernoulli Distribution
- 9:32Beta Distribution
- 15:25Deployment
- 17:02Python Library
related talks
33mTwo Years of Bayesian Bandits for E Commerce
21mLeveraging Multi-Armed Bandit Algorithms for Dynamic Decision Making
41mKeynote: Personalization, Bandits, and Causal inference
32mMulti-Armed Bandits in Recommendations
15mJohn Maxwell, Data Scientist, Nordstrom at MLconf Seattle 2017
30mPractical Multi Armed Bandits
27mEnhancing Personalization Using Recommendation Systems & Multi-Armed Bandits
57mSearch Relevance: A/B testing to Reinforcement Learning