Multi-agent Deep Reinforcement Learning Recommendation Algorithm Based on User Positive and Negative Feedback

Recommendation systems play an increasingly crucial role amidst the prevalence of information overload. However, most existing recommendation systems perceive the recommendation process as static, thus overlooking the implicit information value in interactions between users and systems. Additionally, the majority of current research efforts are focused on processing positive user feedback, while neglecting negative feedback. However, negative feedback also contains valuable user preference information and has the potential to significantly improve recommendation tasks. Therefore, this paper introduces a novel recommendation model based on Deep Reinforcement Learning (DRL), referred to as MDRR-att. In the proposed framework, we presents a specially designed state generation module that incorporates an attention mechanism to extract real-time preference information from users’ historical interaction records. Furthermore, a multi-agent Actor-Critic algorithm is employed to simulate the real-time recommendation process, enabling the dynamic collection of user preference information while effectively utilizing negative feedback data. Experiments were conducted on two publicly available datasets. The results suggest that our proposed recommendation framework offers significant advantages in utilizing user negative feedback information.