Collaborative Information Completion for Robust Graph Collaborative Filtering

Collaborative information learned from the user-item interactions is widely used to present user preferences in recommender systems. Graph collaborative filtering approaches (GCF) could accurately understand and predict the collaborative information from observed interactions by modeling the multi-hop correlations. However, none of the collaborative approaches can get out of the impact of the sparse observation data, leading to unreliable recommendations, especially for users with fewer interactions. To address this problem, we propose a novel approach namely RGCF to enhance GCF via collaborative information completion. According to the impressive performances of recent graph models, we believe that they can effectively model collaborative information from the observed data and accurately predict the user-item correlations. Consequently, RGCF completes the original user-item interactions with the generated user-item correlations to enhance the collaborative information and expand the information propagation path between users and items, thereby improving the performance and accuracy of the recommendation system. Additionally, leveraging a pre-trained model to generate high-quality latent interactions inherently filters out noisy and adversarial signals, ensuring the recommendation system’s robustness and controllability. Such implicit safety mechanisms align with the requirements of secure multimedia analysis, where reliability and stability are critical. We conducted extensive experiments on three public datasets. The results demonstrate that RGCF achieves the best performance in all scenarios. In particular, improvements are achieved by up to 17.2% in sparse settings, verifying the significance of collaborative information completion.