Multi-scale attention graph convolutional network for heterogeneous graph representation

Abstract

Existing heterogeneous graph neural network models focus on improving the efficiency of meta-path aggregation or extending GNN based on node and relationship type parameters, but they overlook the differences in attending to the hierarchical structural information of the graph. We adapt the multi-scale structure originally used for semantic segmentation in images to graph structures, paying more attention to node features and detailed information at higher scales, and focusing on the topological structure information at lower scales. For this, this paper proposes a heterogeneous graph representation learning method based on multi-scale attention graph convolution, MSA-GCN. We map attributes specific to node types into the same feature space, utilizing the hierarchical structure among different node types and relationship types in the heterogeneous graph to explore potential feature correlations. We decouple the heterogeneous graph based on relationship types and partition scales on each subgraph according to the length of meta-paths, aggregating heterogeneous information. Additionally, we apply multi-head attention to each scale of each subgraph, adaptively exploring the importance of the output at each scale in the process of heterogeneous graph representation learning based on the task and dataset. This allows for a more refined and comprehensive aggregation of heterogeneous features and structural information at different levels, enhancing the expressive power of node embeddings. We compared our model with state-of-the-art heterogeneous graph representation models on three different real-world datasets, and the results demonstrated that our model significantly outperformed the best baseline. The maximum performance improvement achieved was 14.81%.