A deep dive into reinforcement learning approaches for recommendation systems and search, covering multi-armed bandits, contextual bandits, value-based methods (DQN), and policy-based methods (REINFORCE). Real-world examples from Yahoo (news article recommendations), Netflix (personalized movie artwork), JD (e-commerce with negative feedback), Microsoft (news with long-term user activeness), ByteDance (joint recommendations and ads), and Google/YouTube (REINFORCE with Boltzmann exploration) are analyzed. Key themes include optimizing for long-term rewards, balancing exploration vs. exploitation, handling cold-start, and off-policy evaluation challenges.
Table of contents
Contextual bandits: Multi-armed bandits with contextValue-based: Learning value for each state-action pairPolicy-based: Learning actions for each state directlyActor-Critic: Combining value-based and policy-basedConclusionReferences1 Impression