Graph Convolutional Networks (GCNs) are network architectures that operate on graph data. Existing GCNs often assume homogeneous graphs which cannot capture the rich semantics of the data, leading to unsatisfactory performance. Many datasets can be more naturally modeled as heterogeneous graphs which reflect intuitively and explicitly the rich semantical information between nodes. There has been little work on designing a GCN on such graph.

We propose AHEG, an Attention-Based HEterogeneous Graph Convolutional Network. As compared with previous work, AHEG retrieves multiple kinds of relationships between different nodes with an efficient meta-path generation mechanism. Furthermore, with a two-stage attention-based convolution to form node embeddings, it assigns values according...[ Read more ]

Graph Convolutional Networks (GCNs) are network architectures that operate on graph data. Existing GCNs often assume homogeneous graphs which cannot capture the rich semantics of the data, leading to unsatisfactory performance. Many datasets can be more naturally modeled as heterogeneous graphs which reflect intuitively and explicitly the rich semantical information between nodes. There has been little work on designing a GCN on such graph.

We propose AHEG, an Attention-Based HEterogeneous Graph Convolutional Network. As compared with previous work, AHEG retrieves multiple kinds of relationships between different nodes with an efficient meta-path generation mechanism. Furthermore, with a two-stage attention-based convolution to form node embeddings, it assigns values according to their importance. To tackle graph-level learning tasks, AHEG has an optional pooling layer to downsample the features while preserving structural information. We conduct extensive experimental study using two transductive graph datasets (DBLP and ACM) and two inductive dataset (PPI and MUTAG). AHEG is shown to substantially outperform the state-of-the-art schemes in terms of node-level or graph-level classification accuracy. Furthermore, it achieves much better NMI/ARI values in clustering analysis.