Higher-Order Heterogeneous Graph Convolutional Network Based on Meta-Paths

Wanting Zhao, Hao Xu, Wenzhuo Huang, Jinkui Xie

Abstract: Graph convolution networks are potent methods in graph representation learning. Meta-paths, which connect different types of nodes, are extensively used to represent various se-mantic meanings in heterogeneous graphs. Inspired by the above, we design a higher-order heterogeneous graph convolutional network based on meta-paths. It not only chooses a few meta-paths but also captures higher-order meta-paths with important higher-order relations (such as communal relation). Besides, it contains a calculation method of higher-order meta-path-based adjacency matrices and a novel heterogeneous graph convolution network for generating node embeddings. At every message passing step, it linearly aggregates information from higher-order meta-path-based neighbors. The computational complexity analysis shows that our proposed model is of high efficiency and applies to large-scale heterogeneous graphs. Our proposed model outperforms the state-of-the-art results in three real-world heterogeneous graphs: DBLP, IMDB, and Amazon Kindle Review. The classification experiments show that the calculation of higher-order meta-path-based adjacency matrices brings 2.23% average accuracy improvement in DBLP and IMDB.