Social-network-based recommendation algorithms leverage rich social network information to alleviate the problem of data sparsity and boost the recommendation performance. However, traditional social-network-based recommendation algorithms ignore high-order collaborative signals or only consider the first-order collaborative signal when learning users' and items' latent representations, resulting in suboptimal recommendation performance. In this paper, we propose a graph neural network (GNN)-based social recommendation model that utilizes the GNN framework to capture high-order collaborative signals in the process of learning the latent representations of users and items. Specifically, we formulate the representations of entities, i.e., users and items, by stacking multiple embedding propagation layers to recursively aggregate multi-hop neighborhood information on both the user-item interaction graph and the social network graph. Hence, the collaborative signals hidden in both the user-item interaction graph and the social network graph are explicitly injected into the final representations of entities. Moreover, we ease the training process of the proposed GNN-based social recommendation model and alleviate overfitting by adopting a lightweight GNN framework that only retains the neighborhood aggregation component and abandons the feature transformation and nonlinear activation components. The experimental results on two real-world datasets show that our proposed GNN-based social recommendation method outperforms the state-of-the-art recommendation algorithms.
Keywords: graph neural network; high-order neighbors; recommendation algorithm; social network.