Decoupled Graph Signal Processing for Noise-Robust Recommendation in Sparse Heterogeneous Networks

Abstract

Real-world recommendation systems usually utilise heterogeneous user-item interaction graphs that are extremely sparse, have multi-type edges and are with considerable observation noise due to misclicks, accidental exposures and bot activities. Traditional graph convolutional recommenders propagate signals indiscriminately across all edges, which amplifies noises and degrades the representation quality, especially for low-degree nodes in sparse areas. We propose DGSP-NR, a Decoupled Graph Signal Processing framework for Noise-Robust Recommendation in Sparse Heterogeneous Networks. DGSP-NR separates the graph propagation operator into a low-frequency signal component and a noise-isolation residual through a learnable spectral decomposition, and then uses type-specific aggregation kernels for heterogeneous edge types. On three sparse benchmarks, Amazon-Books, Yelp-2023 and MIND News, DGSP-NR consistently improves the state-of-the-art graph-based recommenders by up to 8.2% Recall@20 and 7.4% NDCG@20, with significant improvements of 14.1% on ultra-sparse user subsets.