<p>Social recommendation systems have significant potential to enhance recommendation performance by integrating users’ social relationships and user-item interaction information. However, in practical applications, they are often affected by noise, which primarily manifests in two forms: low homophily in social relationships and user rating bias. Low homophily refers to the phenomenon where, despite existing social connections between users, their interests, preferences, and behaviors may differ significantly. Rating bias occurs when certain users consistently give low ratings, which may not accurately reflect their negative attitudes toward the items. Such noise can undermine the model’s ability to accurately capture users’ true preferences, thereby reducing recommendation performance. The core challenge lies in effectively distinguishing valuable signals from noisy information and accurately modeling complex user-item relationships in noisy environments to ensure recommendation reliability and accuracy. To address these challenges, we propose a noise-tolerant social recommendation model, termed GNN_MAM, which is built upon graph neural networks and incorporates an improved multi-head attention mechanism. First, a consistency mask is designed to dynamically adjust the weights of users’ neighbors by evaluating consistency scores, retaining only highly consistent neighbors to mitigate interference from low homophily. Second, a multi-head attention mechanism based on rating bias is introduced to enhance the model’s ability to perceive rating bias and to better capture users’ true preferences. For optimization, a joint learning framework combines supervised loss with self-supervised contrastive learning, generating augmented views to capture more collaborative signals. Experiments on three public datasets—Ciao, Epinions, and Yelp demonstrate that the proposed method achieves superior performance.</p>