In the field of social recommendation, due to the typical graph structure of recommendation and social data, and the relative sparsity of recommendation data, a representative approach has emerged which integrates graph convolutional networks with self-supervised enhancement. However, existing methods based on this framework have shortcomings. Firstly, the message passing mechanism of graph convolutional networks struggles to effectively filter noise, and current designs fail to fully extract implicit cross-graph information. Secondly, existing self-supervised enhancement designs are relatively heavy and may exhibit potential performance instability. To address these issues, we propose the "Representation Contrastive Learning and Distillation Enhanced Denoised Spectral Graph Convolution Network for Social Recommendation (RDDGCN)". This model deeply analyzes the inherent flaws of non-spectral graph convolutions in denoising and devises a spectral-feature-based denoised graph convolution network for generating user/item node representations. Leveraging denoised user/item node representations, we design a knowledge distillation-enhanced social recommendation task to thoroughly extract spectral-domain cross-graph information, effectively boosting the recommendation task. To efficiently extract spectral-domain self-supervised signals, we introduce a lightweight and stable representation-enhanced contrastive learning task as a self-supervised auxiliary task, jointly trained with the knowledge distillation-enhanced social recommendation task. Extensive experiments on real-world datasets demonstrate the superiority of our approach over competing methods. Moreover, component analysis and ablation experiments confirm the rationality of the denoised spectral graph convolution network and downstream task design.
Feng et al. (Mon,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: