Heterogeneous graph neural networks analysis: a survey of techniques, evaluations and applications

Graph Neural Networks (GNNs) have achieved excellent performance of graph representation learning and attracted plenty of attentions in recent years. Most of GNNs aim to learn embedding vectors of the homogeneous graph which only contains single type of nodes and edges. However, the entities and their interactions in real world always have multiple types and naturally form the heterogeneous graph with rich structural and semantic information. As a result of this, it is beneficial to advance heterogeneous graph representation learning that can effectively promote the performance of complex network analysis. Existing survey papers of heterogeneous graph representation learning summarize all possible embedding techniques for graphs and make insufficient analysis for deep neural network models. To tackle this issue, in this paper, we systematically summarize and analyze existing heterogeneous graph neural networks (HGNNs) and categorize them based on their neural network architecture. Meanwhile, we collect commonly used heterogeneous graph datasets and summarize their statistical information. In addition, we compare the performances between HGNNs and shallow embedding models to show the powerful feature learning ability of HGNNs. Finally, we conclude the application scenarios of HGNNs and some possible future research directions. We hope that this paper can provide a useful framework for researchers who interested in HGNNs.

Automated summary

Graph Neural Networks (GNNs) have achieved excellent performance in graph representation learning and have attracted plenty of attention. Most GNNs focus on learning embedding vectors of homogeneous graphs, but in real-world scenarios, entities and their interactions often form heterogeneous graphs with rich information. Therefore, advancing heterogeneous graph representation learning is beneficial for complex network analysis performance.