Relation-aware Diffusion-Asymmetric Graph Contrastive Learning for Recommendation

ACL ARR 2025 May Submission81 Authors

Abstract:

Collaborative filtering (CF) based recommendation has been significantly enhanced by Graph Neural Networks (GNNs) and Graph Contrastive Learning (GCL), yet two persistent challenges remain: (i) random edge perturbations often destroy vital structural signals, degrading semantic consistency across augmented views; and (ii) data sparsity undermines generalization by limiting the propagation of collaborative signals. To address these issues, we propose $$\textbf{R}elation-\textbf{a}ware \textbf{D}iffusion-\textbf{A}symmetric Graph Contrastive Learning for \textbf{R}ecommendation (\textbf{RaDAR}) $$, a novel contrastive framework that integrates two complementary view generation strategies: a graph generative model and a relation-aware graph denoising model. RaDAR introduces three key innovations: (1)asymmetric contrastive learning with global negative sampling to preserve semantic consistency while reducing noise; (2)diffusion-guided augmentation, which improves robustness through progressive noise injection and denoising; and (3)relation-aware edge refinement, which dynamically adjusts edge weights based on latent node semantics. Extensive experiments on three public benchmarks show that RaDAR consistently outperforms state-of-the-art recommendation methods, especially under noisy and sparse settings.
The code of our method is available at our repository(https://anonymous.4open.science/r/RadarGCL-DB7B).

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: Recommendation, Diffusion Model, Contrastive Learning, Data Augmentation

Languages Studied: English

Submission Number: 81