Robust enhanced collaborative filtering without explicit noise filtering

Graph convolutional neural networks have been successfully applied to collaborative filtering to capture high-quality user-item representations. Despite their remarkable performance, there are still limitations that hinder further improvement of recommender systems. Most existing recommendation systems utilize implicit feedback data for model training, but such data inevitably contains adversarial interaction noise. The conventional graph-based collaborative filtering method fails to effectively filter out this noise, and instead amplifies its impact, resulting in degraded model performance. To address this issue, we propose a robustness-enhanced collaborative filtering graph neural network model that does not rely on explicit noise filtering. Our approach involves simulating user-item interactions that do not exist in practice as adversarial interaction noise using random noise. To mitigate the impact of this noise in hidden feedback, we replace them with randomly selected partial nodes based on the principle of mutual information maximization. Our model has been extensively experimented on three public datasets (MovieLens-1 M, Yelp, and Ta-feng) and achieves performance improvements of about 5%, 10%, and 14%, respectively, compared to the state-of-the-art baseline model. In particular, in model robustness experiments, our model achieves significant performance improvements of about 13% and 17% in Yelp and Ta-feng. A comprehensive experimental study shows that our proposed method is reasonably effective and interpretable.