FAHC: frequency adaptive hypergraph constraint for collaborative filtering

Yu Tang, Lilan Peng, Zhendong Wu, Jie Hu, Pengfei Zhang, Hongchun Lu

Abstract: Graph neural networks (GNNs) exhibit superior recommendation performance with their powerful capability of representing complex relationships. However, existing methods encounter two key challenges: (1) The high-frequency signals on graphs are momentous, but the graph convolutional networks cannot adaptively capture the different combinations of various frequency features (i.e., high-frequency and low-frequency signals) simultaneously. (2) GNNs can only integrate the adjacency node features (i.e., pairwise relation), but non-adjacent nodes are also correlated in the user-item interaction graph (i.e., the high-order interaction). To address these challenges, this study explores Frequency Adaptive Hypergraph Constraint for Collaborative Filtering (FAHC). Specifically, FAHC mainly consists of frequency adaptive graph convolutional networks and hypergraph convolutional networks. The frequency adaptive convolutional network can automatically and effectively capture the different combinations of various frequency signals on the graph. Then, we combine the frequency adaptive graph convolutional network with the hypergraph convolutional network to learn the local and global node features. Furthermore, we propose a novel constraint loss, which can help achieve better recommendation performance. The experiments indicate that FAHC improves the other baselines implemented on three published datasets, with the maximum improvement being over 90%. All source codes can be accessed at https://github.com/tangyu-ty/FAHC.