Neural computing and applications

Collaborative filtering Graph neural network Hypergraph learning Recommender systems

In recent years, graph convolution networks and hypergraph convolution networks have become a research hotspot in collaborative filtering (CF) because of their information extraction ability in dealing with the user-item interaction information. In particular, hypergraph can model high-order correlation of users and items to achieve better performance. However, the existing graph-based CF methods for mining interactive information remain incomplete and limit the expressiveness of the model. Moreover, they directly use low-order Chebyshev polynomials to fit the convolution kernel of graph and hypergraph without experimental proof or analysis, lacking interpretability. We propose an effective hybrid graph and hypergraph convolutional network (EHGCN) for CF to obtain a capable and interpretable framework. In EHGCN, the graph and the hypergraph are used to model the correlation among nodes in the interaction graph for multilevel learning. EHGCN also optimizes the information flow framework to match the improved convolution strategy of the graph and hypergraph we proposed. Extensive experiments on four real-world datasets show the considerable improvements of EHGCN over other state-of-the-art methods. Moreover, we analyze the graph and hypergraph convolution kernel in terms of the spectral domain to reveal the core of the graph-based CF, which has a heuristic effect on future work.

Ronghui Guo,Jianwen Chen,Youpeng Hu,Meixia Qu

Xunkai L

Journal paper

Bin Jiang*

Engineering

no

SCI