Dual-level graph contrastive collaborative filtering

The latest research positions graph-based collaborative filtering as an effective strategy in recommendation systems, enabling the analysis of user preferences via user-item interaction graphs. However, such methods often struggle with data sparsity issues in real-world scenarios. To address this, contrastive learning mechanisms have been integrated into graph collaborative filtering, though existing approaches are limited to single-view designs at either the graph or node level, restricting overall model performance. In response, we propose an innovative dual-level graph contrastive collaborative filtering method (DL-GCL) that combines both graph and node-level views. First, at the node level, we employ a matrix decomposition technique during the preprocessing phase to decompose and reconstruct the bipartite graph. Based on the reconstructed results, contrastive views are constructed to capture local collaborative information. Subsequently, considering the potential noise introduced by node-level views, we mitigate the impact of uncertain noise by capturing the model's maximum gradient state at the graph level. Using the Fast Gradient Sign Method (FGSM), we perturb the model's representation vectors under worst-case conditions, thereby mitigating noise from node-level views and extracting global collaborative information. Finally, DL-GCL employs a multi-task learning strategy to optimize local-global views and BPR (Bayesian Personalized Ranking) loss functions. Through extensive experiments on four public datasets, the evaluation metrics NDCG and Recall show up to a 24.5% improvement compared to the latest graph contrastive models. This highlights the strong performance of DL-GCL in improving recommendation system robustness and mitigating data sparsity.