Abstract
Heterogeneous graphs can accurately and effectively model rich semantic information and complex network relationships in the real world. As a deep representation model for nodes, heterogeneous graph neural networks (HGNNs) offer powerful graph data processing capabilities and exhibit outstanding performance in network analysis tasks. However, the traditional HGNNs model cannot mine all the semantic text information during text attribute vector learning; this results in a loss of node information, which degrades the model performance. In this study, we propose a novel framework for the HGNN model enriched with text attribute (HGNN-ETA), which incorporates the text attributes of the nodes. This process involves the attention mechanism-based encoding of the node text attributes, neighborhood aggregation mechanism-based completion of the node attributes, and end-to-end construction and optimization of a heterogeneous graph model. Extensive experiments and comparisons were conducted on two real-world heterogeneous graph-structured datasets. The results demonstrate that the proposed framework is more effective than other state-of-the-art models.