SocialLGN: Light graph convolution network for social recommendation

Abstract

Graph Neural Networks have been applied in recommender systems to learn the representation of users and items from a user-item graph. In the state-of-the-art, there are two major challenges in applying Graph Neural Networks to social recommendation: (i) how to accurately learn the representation of users and items from the user-item interaction graph and social graph, and (ii) based on the fact that each user is represented simultaneously by the two graphs, how to integrate the user representations learned from these two graphs. Aiming at addressing these challenges, this paper proposes a new social recommendation system called SocialLGN. In SocialLGN, the representation of each user and item is propagated in the user-item interaction graph with light graph convolutional layers; in the meantime, the user’s representation is propagated in the social graph. Based on this, a graph fusion operation is designed to aggregate user representations during propagation. The weighted sum is applied to combine the representations learned by each layer. Comprehensive experiments are conducted on two real-world datasets, and the result shows that the proposed SocialLGN outperforms the SOTA method, especially in handling the cold-start problem. Our PyTorch implemented model is available via https://github.com/leo0481/SocialLGN.