How we personalized onet.pl with multi-armed bandits
summary
Personalization on onet.pl uses multi-armed bandits to pick ten articles from hundreds for ~11 million monthly users, with implementations of epsilon‑greedy, UCB and Thompson sampling tested in a simulator across 40 configurations and validated by A/B tests. The pipeline logs clicks as reward, supports pinning sponsored items, and deploys Thompson sampling to production; it improves engagement while handling new items and shifting article life cycles.
AI-written summary · about
chapters (13)
- 0:00Personalized recommendations intro
- 1:56News recommendation problem
- 4:23Multi-armed bandit problem
- 6:52Recommendations as bandit problem
- 7:56Epsilon-greedy bandit solutions
- 10:07UCB bandit for exploration
- 13:00Thompson sampling estimates
- 15:49Bandit algorithms summary
- 16:53Dynamic article popularity
- 19:11Parameter tuning challenges
- 20:19Simulator for testing
- 22:55Onet.pl implementation
- 24:00Q&A
related talks
21mLeveraging Multi-Armed Bandit Algorithms for Dynamic Decision Making
39mGrowing user engagement with RL-driven personalisation
30mPractical Multi Armed Bandits
27mEnhancing Personalization Using Recommendation Systems & Multi-Armed Bandits
33mTwo Years of Bayesian Bandits for E Commerce
32mMulti-Armed Bandits in Recommendations
22mPersonalizing Explainable Recommendations with Multi-objective Contextual Bandits
41mKeynote: Personalization, Bandits, and Causal inference