Embedding Refinement through GAN Framework for Graph Recommender System

The core challenge of recommender systems is how to effectively and comprehensively discover relationships between users and items. In recent years, Graph Neural Networks (GNNs) have provided a new paradigm for recommendation systems by explicitly modeling the relationships of user and item by graph. Their strong capability to capture higher-order relationships has made them widely adopted in recommendation tasks. However, in practice, the accuracy of GNNs based recommender system often suffers from the low quality of the input data and suboptimal constructed graph. In reality, most of the datasets inevitably suffer from data sparsity and noise, which can degrade the quality of the input in-formation and representation ability of the constructed graph. To address this issue, inspired by Generative Adversarial Networks (GANs), we propose a generic embedding refinement method. Our approach enhances embedding representations through adversarial training between the generator and discriminator, thereby improving both the robustness and effectiveness of the model.