Motif-Based Hypergraph Convolution Network for Semi-Supervised Node Classification on Heterogeneous Graph

GNN(Graph neural network)-based representation learning methods have been a subject of intense research activities on network representation learning in recent years because of their significant contribution to the extraction of network structure information. Compared with the widely studied homogeneous information network, however, the networks in the real world are often HIN (heterogeneous information networks) composed of different types of objects interconnected by complex relationships. The complex structural information and rich semantic information bring great challenges for network representation learning for HIN. Exploiting the higher-order structural patterns is an effective way of learning the representation for the complex network, the traditional GCNs are not directly applied to the higher-order heterogeneous network. The motifs have been recognized as the most common way for understanding and exploring the complex networks, which can both describe complex semantic information and preserve high-order neighbor structures in the network. We leverage the high-order network patterns in the form of motifs to transform the heterogeneous network into multiple hypergraphs, and then model multiple hypergraphs to learn node representations in a unified way. Compared with the traditional heterogeneous network, the advantages of a heterogeneous hypergraph network are that the heterogeneous properties can be fully expressed by introducing high-order structural information and fusing the different types of semantic hyperedges based on various motifs. Therefore, in order to overcome the deficiency that homogeneous network can only describe the pairwise relationship, we proposed a heterogeneous MHGCN (Motif-based HyperGraph Convolutional Network) to model the recurring high-order network patterns as hyperedges formed by multiple related nodes, and transform the entire heterogeneous information network into multiple hypergraphs that composed of different hyperedges. In a homogeneous network, the GCN relies on localized first-order approximations of spectral graph convolutions. However, for a heterogeneous network, the convolutional operation on the motif-based hypergraph should be redefined for preserving all the relevant node features. As convolution operations on hypergraph can be conducted using the basic properties of a hypergraph and spectral graph theory, to overcome the issue that the motif-based sampling strategy can not cover all nodes, we integrate multiple hypergraph networks and the original network structural information into hypergraph convolutional network to ensures that all nodes can be aggregated and updated in the process of propagating node features. Inspired by the attention mechanism, we exploit the hypergraph attention network to dynamically aggregate node features based on their importance and semantic roles. Specifically, we conduct two levels of attention networks with a hierarchical structure, namely a hyperedge-level attention to learn the importance among different types of nodes, while comprehensive semantic-level attention to capture the importance of different types of motif structures. Owing to the proposed MHGCN is in an end to end fashion, it can learn to get the labels of the nodes eventually. The proposed MHGCN improves the micro-F1 by 0.56%~3.51%, improves the macro-F1 by 0.54%~4.37% on two real-world datasets of DBLP-P and DBLP-A in semi-supervised node classification task against the state-of-the-art methods, which also verifies the effectiveness of aggregating different types of nodes based on hyperedge-level attention and comprehensive semantic-level attention. © 2021, Science Press. All right reserved.