Signal Contrastive Enhanced Graph Collaborative Filtering for Recommendation

Abstract Graph collaborative filtering methods have shown great performance improvements compared with deep neural network-based models. However, these suffer from data sparsity and noise problems. To address issues, we propose a new contrastive learning-based graph method to learn more robust representations. The proposed is called signal enhanced (SC-GCF), which conducts learning on signals. It has been proved that networks correspond low-pass filters the signals convolution perspective. Different previous methods, first pay attention diversity of directly optimize informativeness We introduce hypergraph module strengthen representation ability networks. utilizes learnable structure model latent global dependency relations cannot depict. Experiments are conducted four public datasets, results show significant state-of-the-art confirms importance considering signal-level learning.