Higher-Order Graph Contrastive Learning for Recommendation

ZhenZhong Zheng, Jianxin Li, Xiaoming Wu, Xiangzhi Liu, Lili Pei

Abstract: Collaborative filtering, as an effective recommendation technique, models user preferences by representing user interactions as a bipartite graph(user-item). However, the graph-based model struggles to mitigate the impact of data sparsity. Recent studies have attempted to tackle this problem by utilizing contrastive learning. Nevertheless, most of these methods rely on augmenting the data based on the original graph to construct contrastive views, neglecting the utilization of higher-order relationships(user-user, item-item). Alternatively, some methods erroneously compare higher-order relationships directly with the original graph. Such an asymmetric contrasting approach makes it challenging to fully exploit higher-order knowledge. To address the issues, we propose HoGCL, a novel contrastive learning paradigm. HoGCL enriches representation learning by mining higher-order dependencies from the original graph to enhance supervisory signals. Specifically, we construct two contrasting views: higher-order and general views. In the higher-order view, we devise a high-order symmetric contrastive scheme to better explore higher-order dependencies. For the general view, the objective is to fully leverage the learned higher-order knowledge and transfer it into node representations. Moreover, to remove redundant information from higher-order knowledge, HoGCL employs an adaptive fusion strategy to obtain the final representations. Experimental results demonstrate the effectiveness of HoGCL, which significantly outperforms existing models..