Enhancing robustness in implicit feedback recommender systems with subgraph contrastive learning

Abstract

Contrastive learning operates by distinguishing differences between various nodes to facilitate item recommendations. However, current graph contrastive learning (GCL) methods suffer from insufficient robustness. To mitigate the impact of noise and accurately capture user preferences, we propose a subgraph-based GCL method: SubGCL. Firstly, we devise a dynamic perceptual signal extractor that leverages node degree and neighborhood information to model subgraphs corresponding to nodes and compute mutual information scores. This approach enhances view adaptivity, thereby improving data augmentation robustness against noise perturbations. Secondly, we develop an association graph self-attention propagation mechanism. This mechanism constructs node clusters by randomly sampling nodes and edges, facilitating self-attention propagation on the graph to learn cluster associations and enhance recommendation accuracy. Finally, we reconstruct graph structures through recommendation loss and update node embeddings via contrastive learning to bolster the model’s accuracy and robustness in implicit feedback data. We conducted experiments on three publicly available real-world datasets. Results demonstrate that, compared to existing contrastive learning recommendation approaches, SubGCL achieves an average improvement of 4.96% and 3.98% in Recall and NDCG metrics, respectively.