Rule-guided Skip-GCN in neural latent information diffusion network for social recommendation

Graph neural networks (GNNs) have significantly advanced social recommendation; however, existing methods often homogenize diverse social influences and neglect latent item-item correlations, limiting their ability to model complex user preferences. To overcome these limitations, we propose DiffRSG, a novel framework that integrates neural latent information diffusion with explicit rule-guided reasoning. Specifically, DiffRSG employs a rule-guided graph convolutional network (GCN) with skip connections, termed Skip-GCN, to capture heterogeneous social interactions and uncover implicit item dependencies by using explicit rules, and incorporates a specialized prediction layer for accurate rating estimation. Extensive experiments conducted on the Yelp and Flickr datasets demonstrate that DiffRSG significantly outperforms eleven baseline models in terms of hit ratio (HR) and normalized discounted cumulative gain (NDCG), achieving an average relative improvement of over 15% compared to the strongest baseline, DiffNet++, on the primary HR@10 and NDCG@10 metrics. These results demonstrate the effectiveness of differentiating social tie strengths and capturing latent item correlations within a unified framework.