Reinforcement learning-based denoising network for sequential recommendation

Sequential recommendation models each user as a chronological sequence of interacted items and aims to predict what the user will buy in the near future. In this task, sequential dependency is an important factor that needs to be considered, as items the user will buy in the future are largely dependent on past items. Due to the randomness and diversity of user behaviors, not all items purchased are relevant to the next choice. Considering that these irrelevant items tend to bury the influence of the few truly relevant items, how to extract reliable items relevant to the target item to make a correct recommendation is thus a crucial and challenging task. In this paper, we propose a R einforcement L E arning-based D enoising network(RED for short) to make data denoising for sequential recommendation. Specifically, RED formalizes the sequential denoising problem into a Markov Decision Process and utilizes a reinforcement learning method to automatically separate the initial sequence into two parts: one contains items relevant to the target item, while the other one keeps irrelevant items. A Pseudo-Siamese component is further applied to these two subsequences to generate their delayed rewards, and a pairwise policy gradient strategy is designed to guarantee the robustness of the learning process. The experiments on four public datasets demonstrate that our model outperforms state-of-the-art recommendation methods on a variety of common evaluation metrics.