Graph Neural Network with Heterogeneous Attributes for Node Classification

Node classification on heterogeneous graphs with multiple node types and relationships has been widely used in commodity classification, friend recommendation, and so on. Most of the existing models need to define the meta paths manually with incomplete attribute information, leading to the low accuracy of classification. To improve the accuracy of node classification task, we propose a novel graph neural network with heterogeneous attributes (GNNHA) by introducing the heterogeneous attributes of nodes and edges without defining meta paths. First, we design a spatial mapping for heterogeneous edges by mapping and fusing the source and target nodes to a hidden layer space. Then, we make a unique linear projection for each node to obtain the embeddings containing the attributes of nodes and edges by aggregating the information of neighbor nodes. Finally, we provide an improved loss function by leveraging the information loss of mapping and node class imbalance. Our model can capture more complete heterogeneous graph attributes and obtain more objective and accurate classification results. Experimental results show that our proposed model outperforms state-of-the-art ones.