Abstract
<title>Abstract</title> <p>Graph neural networks (GNNs) have achieved promising results in recommender systems, yet their performance often degrades in the presence of noisy implicit feedback and sparse interactions. Existing approaches commonly represent the user-item graph as a single structure, which limits their ability to distinguish stable collaborative patterns from unreliable local connections. This paper presents a Contrastive Multi-Decomposition Approach with Social-Guided Signals (COMPASS), which combines multi-resolution low-rank decomposition with social-guided contrastive learning. Instead of relying on stochastic graph augmentations, COMPASS constructs multiple views from rank-truncated singular value decomposition , where lower-rank components capture global collaborative information and higher-rank components preserve finer interaction details. These views are integrated through a Gumbel-Softmax gating mechanism that learns the relative importance of different approximation levels. Social relations are incorporated as a soft contrastive regularizer to encourage socially consistent representations. In addition, an adaptive gradient-balanced weighting strategy is adopted to improve the stability of multi-objective optimization. Experiments conducted on Yelp, LastFM, and Douban show that COMPASS consistently achieves competitive improvements over strong baselines, with gains of up to 5.76% in NDCG@20. Additional noise injection experiments indicate that the proposed framework maintains stable performance under corrupted interactions, suggesting favorable robustness to noisy feedback.</p>