Denoising social recommendation with self-supervised learning

Abstract

Social recommendation is increasingly prominent in online platforms such as e-commerce and streaming services, where integrating social data can enhance the modeling of user-item interactions. By leveraging the social connections between users, platforms can better predict preferences and provide more personalized recommendations. However, traditional recommendation systems often struggle with data sparsity, which limits their effectiveness. To address this challenge, recent advancements in Self-Supervised Learning (SSL) has shown promising results. SSL mitigates data sparsity issues by utilizing auxiliary learning tasks. Building on this success, scientists have begun to integrate SSL with collaborative social filtering. This integration augments the main recommendation objective with auxiliary task that are aware of social context. Despite these advances, a significant challenge remains: social information often includes noise that can distort the accurate representation of user preferences, especially due to social connections unrelated to users’ interests, like those with colleagues or classmates. To overcome this problem, we introduce a new model named the Denoised Self-Augmented Learning paradigm (DSL). This paradigm retains beneficial user-user connections to improve user behavior learning but also facilitates cross-aspect information sharing through adaptive content matching in the latent space. The experiments on different real-world datasets show that our model surpasses current state-of-the-art model.