A Stable Deep Reinforcement Learning Framework for Recommendation

Recommender system (RS) solves the problem of information overload, which is crucial in industrial fields. Recently, reinforcement learning (RL) combined with RS has attracted researchers' attention. These new methods model the interaction between RS and users as a process of serialization decision-making. However, these studies suffer from several disadvantages: 1) they fail to model the accumulated long-term interest tied to high reward, and 2) these algorithms need a lot of interactive data to learn a good strategy and are unstable in the scenario of recommendation. In this article, we propose a stable reinforcement learning framework for recommendation. We redefine the Markov decision process of RL-based recommendation, and add a stable module to model high feedback behavior of users. Second, an advanced RL algorithm is introduced to ensure stability and exploratory. The experiments verify the effectiveness of the proposed algorithm.

Automated summary

This article proposes a stable reinforcement learning framework for recommendation systems, addressing previous research limitations by redefining the Markov decision process of RL and introducing a stable module. The experiments confirm the effectiveness of the proposed algorithm.