Improving Collaborative Filtering Recommendation via Graph Signal Processing

Recommender systems provide personalized predictions by identifying items likely to interest each user. $k$-NN-based collaborative filtering (CF) is a widely used recommendation strategy, but it typically relies on dense graphs and a fixed neighborhood size, leading to unnecessary computational cost and suboptimal structure modeling. In this paper, we leverage graph signal processing (GSP) to learn a sparse, high-quality user graph that accelerates CF while achieving even better accuracy. Experiments on benchmark datasets show very promising results.