A Multi-Agent Reinforcement Learning Framework for Personalized Product Recommendation in Live-Streaming E-Commerce: Integrating Improved QMIX with Graph Attention Networks and Deep Reinforcement Learning

Abstract

With the rapid development of live-streaming e-commerce, achieving accurate and personalized product recommendations in a dynamic, real-time, multi-user environment poses a serious challenge to traditional methods. These methods struggle to effectively integrate global collaborative signals and accurately model the dynamic evolution of user interests. This study proposes a personalized recommendation optimization framework that combines a Q-value mixing network algorithm for multi-agent reinforcement learning (QMIX). The framework introduces a multi-user information fusion module built on a graph attention network (GAT) and recurrent neural network (RNN), which constructs unified team-level state representations from heterogeneous local agent observations. Building on this fused representation, a Soft Actor-Critic (SAC)-based deep reinforcement learning mechanism with dual gate recurrent unit encoders models users' short-term interaction dynamics and long-term preference stability, enabling continuous strategy optimization. Experiments on the Movi-elens-100K and Amazon datasets demonstrate that the proposed model achieves Accuracy@20 values of 0.2051 and 0.0803, respectively, outperforming strong baselines including OneRec and DSETRec across Accuracy@20, Recall@20, F1@20, and cumulative reward metrics. The proposed framework offers a scalable and effective solution for multi-user collaborative recommendation in live-streaming commerce scenarios.