Meta-path convolution based heterogeneous graph neural network algorithm

Abstract

In the multilayer graph convolution calculation, each node is usually represented as a single vector, which makes the high-order graph convolution layer unable to distinguish the information of different relationships and sequences, resulting in the loss of information in the transmission process. To solve this problem, a heterogeneous graph neural network algorithm based on meta-path convolution was proposed. Firstly, the feature transformation was used to adaptively adjust the node features. Secondly, the high-order indirect relationship between the nodes was mined by convolution within the meta-path to capture the interaction between the target node and other types of nodes under the element path. Finally, the reciprocity between semantics was explored through the self-attention mechanism, and the features from different meta-paths were fused. Extensive experiments were carried out on ACM, IMDB and DBLP datasets, and compared with the current mainstream algorithms. The experimental results show that the average increase of Macro-F1 in the node classification task is 0.5%～3.5%, and the ARI value in the node clustering task is increased by 1%～3%, which proves that the algorithm is effective and feasible.

Keywords

• heterogeneous graph;graph embedding;graph neural network;meta-path;graph convolution