Graph Neural Network (GNN)-based social recommendation algorithms leverage social relations to improve recommendation performance. However, directly incorporating raw social graph often introduces noise due to non-homophilous connections. Moreover, existing negative sampling strategies typically adopt a fixed hardness, which may generate false negatives and weaken the model’s ability to capture user preferences. To address these limitations, this paper proposes a recommendation algorithm based on social diffusion and adaptive negative sampling. Specifically, we first apply forward diffusion and interest-guided denoising on the social network to obtain homophilic social representations. Then, a multi-view alignment mechanism is introduced to maximize mutual information among user representations from the denoised social graph, the raw social graph, and the user-item interaction graph, thereby enhancing representation quality. In addition, adaptive negative sampling dynamically adjusts sample hardness according to the predicted scores of positive items, calibrating the decision boundary between positive and negative instances. Extensive experiments demonstrate that the proposed model consistently outperforms other advanced social recommendation approaches. On the Douban dataset, Recall and NDCG increase by 12% and 10.5% respectively; on Epinions, by 15.6% and 11.1%; and on Yelp, by 13.8% and 14.9%, highlighting its effectiveness in mitigating noise interference and distinguishing subtle differences between positive and negative samples.