Dual-Space Gated Fusion Adaptive Denoising and Diversity-Balanced Recommendation

Abstract

Graph neural networks have become a widely adopted backbone for collaborative filtering because they capture high-order dependencies in user–item interaction graphs. However, noisy interactions can be amplified during graph propagation, while conventional contrastive objectives may over-separate behaviorally similar users and thereby compromise the balance between accuracy and diversity. To address these issues, we propose GAD2-DRec, a Dual-Space Gated Fusion Adaptive Denoising and Diversity-Balanced Recommendation model. The model fuses attribute- and topology-space representations through a gated unit, suppresses unreliable edges with a learnable confidence threshold, and uses a Jaccard-weighted contrastive objective with candidate-based negative sampling to moderate negative-pair repulsion. Experiments on the Amazon and Yelp datasets show that the proposed model improves recommendation quality and coverage while maintaining stronger robustness under randomly injected interaction noise. The code is available at https://github.com/aruivvd/-GAD2-DRec