MRAN: a attention-based approach for social recommendation

Graph Neural Networks have been widely used in social recommendation systems. However, with the increase of graph nodes and diffusion depth, they tend to suffer from graph sparsity and over-smoothing, which inhibit their performance. In this work, we propose the multi-relational attention network, named as MRAN, for social recommendation. Our model has three distinctive characteristics: (i) it alleviates the data sparsity problem in social recommendation scenarios by incorporating both user social relations and item homogeneous relations as supplementary information; (ii) it mimics the structure of influence diffusion in user and item domain via an iteratively aggregating structure; (iii) it has a two-level attention mechanism at the diffusion and aggregating level, enabling it to differentiate importance of embeddings to overcome the over-smoothing problem. Experiments conducted on two large-scale representative datasets demonstrate that the proposed model outperforms previous methods substantially. The ablation study shows that the performance of MRAN can be further improved avoid over-smoothing by increasing the diffusion depth.