A Two-Stage Social Recommendation Framework with Dual-Graph Dynamic Denoising and Bidirectional Cross-View Contrastive Learning

Abstract

Social recommendation models leverage graph neural networks (GNNs) to jointly capture user–item interactions and user–user relationships, thereby improving recommendation performance. However, noise in both graphs may accumulate during message passing and disrupt representation learning. Existing methods often denoise only one graph or apply static and uniform criteria to heterogeneous relations, overlooking the distinct noise characteristics of interaction and social graphs. Moreover, structural and semantic discrepancies may remain between the denoised views, hindering cross-view alignment and information fusion. To overcome these challenges, we propose DGDC-SR, a framework that integrates dual-graph dynamic denoising with bidirectional cross-view contrastive learning. During pre-training, we design separate denoising strategies for the two graphs. For the interaction graph, an implicit gradient-based method adaptively estimates the contributions of observed interactions. For the social graph, a multi-signal dynamic soft denoising strategy estimates edge reliability and suppresses noise propagation. During fine-tuning, bidirectional cross-view contrastive learning aligns user representations across the denoised views, while an adaptive gating fusion mechanism effectively integrates user representations from both views. Extensive evaluations on three public datasets demonstrate the effectiveness of DGDC-SR, which consistently outperforms representative baseline methods.