Discovering Localized Information for Heterogeneous Graph Node Representation Learning

Abstract

Representation learning for heterogeneous graphs aims at learning meaningful node (or edge) representations to facilitate downstream tasks such as node classification, node clustering, and link prediction. While graph neural networks (GNNs) have recently proven to be effective in representation learning, one of the limitations is that most investigations focus on homogeneous graphs. Existing investigations on heterogeneous graphs often make direct use of meta-path type structures. Meta-path-based approaches often require a priori designation of meta-paths based on heuristic foreknowledge regarding the characteristics of heterogeneous graphs under investigation. In this paper, we propose a model without any a priori selection of meta-paths. We utilize locally-sampled (heterogeneous) context graphs “centered” at a target node in order to extract relevant representational information for that target node. To deal with the heterogeneity in the graph, given the different types of nodes, we use different linear transformations to map the features in different domains into a unified feature space. We use the classical Graph Convolution Network (GCN) model as a tool to aggregate node features and then aggregate the context graph feature vectors to produce the target node's feature representation. We evaluate our model on three real-world datasets. The results show that the proposed model has better performance when compared with four baseline models.