On Bandits, Bayes, and swipes: gamification of search
summary
A swipe‑based car search app actively learns a user’s taste from a few gestures by treating features (brand, body, segment) as separate multi-armed bandits and selecting arms with UCB to trade off exploration and exploitation. PCA plus clustering picks ~10 representative cars to avoid popularity bias, Elasticsearch ranks results, and Python (numpy scipy sklearn flask) wires the pipeline. Claim: small data yields good personalization with a few dozen swipes; it ships and scales.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 0:49Gamification to active learning
- 3:59What is active learning
- 7:29Defining interestingness
- 10:18Active learning beyond classification
- 12:29Multi-armed bandits
- 14:15UCB strategy in bandits
- 17:33Live demo: UCB algorithm
- 22:20Scaling bandits to large spaces
- 25:14Handling data bias
- 27:01System architecture and tools
- 28:09Key takeaways
- 29:33Q&A
related talks
17mByron Galbraith, Chief Data Scientist, Talla, NYC 2017
15mAI-Driven Application Optimization
57mSearch Relevance: A/B testing to Reinforcement Learning
20mWinning the Game: Leveraging Contextual Bandits for Creative Optimization
27mEnhancing Personalization Using Recommendation Systems & Multi-Armed Bandits
27mHow we personalized onet.pl with multi-armed bandits
39mGrowing user engagement with RL-driven personalisation
27mA smooth ride: Online car buying and selling at mobile.de