Thompson Sampling with Time-Varying Reward for Contextual Bandits

Contextual bandits efficiently solve the exploration and exploitation (EE) problem in online recommendation tasks. Most existing contextual bandit algorithms utilize a fixed reward mechanism, which makes it difficult to accurately capture the preference changes of users in non-stationary environments, thus affecting recommendation performance. In this paper, we formalize the online recommendation task as a contextual bandit problem and propose a Thompson sampling algorithm with time-varying reward (TV-TS) that captures user preference changes from three perspectives: (1) forgetting past preferences based on a functional decay method while capturing possible periodic demands, (2) mining fine-grained preference changes from multi-behavioral implicit feedback, and (3) iterating the reward weights adaptively. We also provide theoretical regret analysis to demonstrate the sublinearity of the algorithm. Extensive empirical experiments on two real-world datasets show that our proposed algorithm outperforms state-of-the-art time-varying bandit algorithms. Furthermore, the designed reward mechanism can be flexibly configured to other bandit algorithms to improve them.