Multi-Agent RL-based Information Selection Model for Sequential Recommendation

Kaiyuan Li, Pengfei Wang, Chenliang Li

Abstract: For sequential recommender, the coarse-grained yet sparse sequential signals mined from massive user-item interactions have become the bottleneck to further improve the recommendation performance. To alleviate the spareness problem, exploiting auxiliary semantic features (\eg textual descriptions, visual images and knowledge graph) to enrich contextual information then turns into a mainstream methodology. Though effective, we argue that these different heterogeneous features certainly include much noise which may overwhelm the valuable sequential signals, and therefore easily reach the phenomenon of negative collaboration (ie 1 + 1 > 2). How to design a flexible strategy to select proper auxiliary information and alleviate the negative collaboration towards a better recommendation is still an interesting and open question. Unfortunately, few works have addressed this challenge in sequential recommendation. In this paper, we introduce a Multi-Agent RL-based Information S election Model (named MARIS) to explore an effective collaboration between different kinds of auxiliary information and sequential signals in an automatic way. Specifically, MARIS formalizes the auxiliary feature selection as a cooperative Multi-agent Markov Decision Process. For each auxiliary feature type, MARIS resorts to using an agent to determine whether a specific kind of auxiliary feature should be imported to achieve a positive collaboration. In between, a QMIX network is utilized to cooperate their joint selection actions and produce an episode corresponding an effective combination of different auxiliary features for the whole historical sequence. Considering the lack of supervised selection signals, we further devise a novel reward-guided sampling strategy to leverage exploitation and exploration scheme for episode sampling. By preserving them in a replay buffer, MARIS learns the action-value function and the reward alternatively for optimization. Extensive experiments on four real-world datasets demonstrate that our model obtains significant performance improvement over up-to-date state-of-the-art recommendation models.