PGA-DRL: Progressive graph attention-based deep reinforcement learning for recommender systems

Abstract

Advanced graph models, including Graph Convolutional Networks (GCNs) and Graph Attention Networks (GATs), have demonstrated their effectiveness in capturing intricate user-item interactions. However, their integration into Deep Reinforcement Learning (DRL)-based Recommender Systems (RSs) remains relatively underexplored. To address this gap, we propose PGA-DRL, a Progressive Graph Attention-Based DRL model that incrementally fuses GCN and GAT representations via concatenation, effectively combining their complementary strengths to enhance feature representation within an Actor-Critic (AC) framework. This progressive integration refines both global and localized user-item interaction patterns, Specifically, global patterns capture broader user preferences across the entire graph, and localized patterns focus on specific, detailed interactions between closely connected nodes, enabling a more comprehensive understanding of the recommendation environment. We evaluate our approach using extensive experiments on multiple benchmark datasets, including ML-100K, ML-1M, Amazon Subscription Boxes, Amazon Magazine Subscriptions, and ModCloth, employing standard ranking metrics such as Precision@10, Recall@10, NDCG@10, MRR@10, and Hit@10. The experimental results reveal that PGA-DRL outperforms state-of-the-art baselines, such as BPR, NeuMF, and SimGCL, achieving improvements in NDCG@10 and Recall@10. Our core contributions lie in bridging graph-based learning with reinforcement learning through a novel, efficient, and scalable fusion mechanism that enhances recommendation accuracy and ultimately improves user satisfaction. The source code for PGA-DRL is publicly available at https://github.com/RS-Research/PGA-DRL to enhance transparency and facilitate future research.