Research on node classification of heterogeneous graphs with converged attention

As the data in today's information age grows rapidly, the complexity of information types also increases. From the perspective of graph data, nodes on the graph are no longer limited to a single category. Instead, they are evolving towards multiple categories of nodes and edges of different types. However, most traditional graph neural networks are designed for node classification tasks on homogeneous graphs. There is relatively less research on extracting and classifying target node information on heterogeneous graphs. Among the limited models for heterogeneous graph processing, this study chooses the Heterogeneous Graph Attention Network (HAN) model for the classification of nodes in a heterogeneous graph. Building upon the HAN model, this paper addresses the issues of simple concatenation and gradient explosion by introducing attention mechanisms and optimizing activation functions. The proposed approach achieves significant results in node classification tasks on the publicly available ACM and DBLP datasets, with accuracy reaching 89.55 and 93.66, respectively.