Social-aware graph contrastive learning for recommender systems

Recommender systems usually encounter the issue of sparse interaction data, which is commonly alleviated by social recommendation models based on graph neural networks. However, these models overlook the collaborative similarity relationship among items and fail to effectively integrate and process various graph structures. To address these issues, we propose a novel S ocial -aware G raph C ontrastive L earning R ecommendation model (SG-CLR). Specifically, we initially utilize data augmentation techniques to obtain different augmented views of user-item interaction. Secondly, a social -aware encoder is put forward to effectively capture both the influence diffusing within the social network and the attractiveness of items among the item collaborative similarity graph. Finally, we employ graph contrastive learning to maximize the consistency of node representation across different augmented views, and further focus on domain -shared information through joint training. Experimental results conducted on two real -world datasets demonstrate that the proposed SG-CLR outperforms the state-of-the-art baselines. Compared to the best baseline, SG-CLR improves the performance on the two datasets by 3.069% and 2.972%, respectively.