JLR-GCN: Joint label-aware and relation-aware graph convolution neural network for heterogeneous graph representations

Abstract

Node representation learning on heterogeneous graphs has garnered increasing attention. However, most existing heterogeneous graph learning models primarily focus on improving the performance of neural networks while overlooking two crucial aspects. On the one hand, they do not give sufficient importance to the value of label information in heterogeneous graph representation learning and merely use it for loss computation. On the other hand, they neglect the differences in heterogeneous structural information carried by interactions between nodes of different relation types. To address these challenges, this paper proposes a joint label-aware and relation-aware graph convolution neural network (JLR-GCN) for heterogeneous graph representations. Our model enhances the expressive power of the heterogeneous graph by leveraging label attributes and structural information. Label-aware and relation-aware are combined to explore the differences in the influence of each relation type during the heterogeneous graph representation learning process. Moreover, a stacked graph convolutional network with meta-path importance parameters is adopted for heterogeneous message passing to capture long and short meta-paths. The allocation of network layer outputs based on their importance is adjusted, enabling the effective learning of node embeddings by incorporating valuable heterogeneous relation, node attributes, and structural signals. 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 16.14%.