Treffer: Heterogeneous Multilayer Graph Convolutional Network

Heterogeneous graph convolutional networks have used in various network analytical tasks on heterogeneous network data, like link prediction and node classification. In recent, some graph convolutional networks have been proposed for multiplex heterogeneous graphs. However, there are a few model proposed for both heterogeneous and multiplex heterogeneous graphs. Most of the proposed heterogeneous graph neural networks are based on choosing meta-paths. When the heterogeneous graphs have a lot of node types and edge types, choosing meta-paths is difficult. To tackle these challenges, this work proposes a Heterogeneous Multilayer Graph Convolutional Network (HMGCN) for both heterogeneous network and multiplex heterogeneous graphs embedding. Our HMGCN looks at heterogeneous networks as multilayer networks, which means each edge type shows one layer. Specifically, HMGCN employs two major components, i.e., graph convolutional neural network for each edge type layers, and fusing embeddings of each node types from multiple different layers. In addition, we effectively learn node embeddings by integrating multi nodes types and multi relations types’ structure and attribute semantics into with both unsupervised and semi-supervised learning paradigms. Experiments results on five real datasets with several network analytical tasks show the outstanding superiority of HMGCN against state-of-the-art embedding baselines in terms of all evaluation metrics.