Light disentangled graph learning for social recommendation

Graph Neural Networks (GNNs) have been utilized in social recommendation, leveraging social relations to enhance the representation of learning for recommendation. Most social recommendation models unify the user-item interactions and social relations in the user representation. Although existing recommender systems have made great progress, most methods struggle to effectively capture the diverse behavioral patterns of users across the two domains, and this limitation hampers the ability to represent users and their preferences accurately. To overcome this limitation, we introduce a novel social recommendation disentangled learning framework (LDGSR). Our model not only highlights the significance of incorporating heterogeneous relationships and latent factor decomposition in social network recommendation models but also explores the rich relationships between items. We conducted comprehensive experiments on four widely used benchmark datasets to validate the effectiveness of the proposed method.