融合自适应图生成与去噪机制的社交推荐算法研究

Research on Social Recommendation Algorithm Integrating Adaptive Graph Generation and Denoising Mechanism

摘要: 本研究旨在优化协同过滤推荐算法在面对数据稀疏性时，引入社交网络所带来的噪声干扰问题。现有的图神经网络社交推荐方法虽然能够利用社交辅助信息，但往往忽略了社交关系中的无效连接(如兴趣不一致的表面社交)，且现有的随机数据增强方法容易破坏图的内在语义结构。为此，本文提出了一种融合自适应图生成与去噪机制的对比学习推荐框架。该框架通过在用户侧与物品侧同时部署自适应增强策略：一方面，利用变分图自编码器(VGAE)重构用户社交视图以补充潜在语义，并结合交互信号引导的去噪模块剔除无效社交边；另一方面，在物品侧对交互图本身进行自适应去噪，以增强物品表示的鲁棒性。通过跨视图对比学习，模型实现了辅助信息与主交互任务的有效对齐。在Yelp、CiaoDVD两个公开数据集上的实验结果表明，本研究提出的模型在Recall@20和NDCG@20指标上均显著优于LightGCN、MF等主流基线模型。进一步的分析显示，该模型在处理高噪声比例及稀疏场景下表现出更强的鲁棒性与泛化能力。

Abstract: This study aims to optimize the issue of noise interference introduced by social networks when collaborative filtering recommendation algorithms encounter data sparsity. Although existing graph neural network-based social recommendation methods can utilize social auxiliary information, they often overlook invalid connections in social relationships (such as superficial social ties with inconsistent interests), and the current random data augmentation methods tend to disrupt the intrinsic semantic structure of the graph. To this end, this paper proposes a contrastive learning recommendation framework that integrates an adaptive graph generation and denoising mechanism. This framework deploys adaptive enhancement strategies on both the user side and the item side: on the one hand, it uses a variational graph autoencoder (VGAE) to reconstruct the user social view to supplement latent semantics and combines a denoising module guided by interaction signals to eliminate invalid social edges; on the other hand, it performs adaptive denoising on the interaction graph itself on the item side to enhance the robustness of item representations. Through cross-view contrastive learning, the model achieves effective alignment between auxiliary information and the main interaction task. Experimental results on the public datasets Yelp and CiaoDVD show that the model proposed in this study significantly outperforms mainstream baseline models such as LightGCN and MF in terms of Recall@20 and NDCG@20 metrics. Further analysis indicates that this model demonstrates stronger robustness and generalization ability in handling scenarios with high noise ratios and sparsity.