Heterogeneous graph neural networks with denoising for graph embeddings

With the increasing popularity of graph structures, Graph embedding, Which aims to project nodes into low dimensional space while preserving the topological structure information of graphs and the information of nodes themselves, Has attracted an increased amount of attention in recent years. most of the embedding methods based on heterogeneous graphs use a meta-path guided random walk to capture the semantic and structural correlation between different types of nodes in the graph. despite the success of the meta-path-guided heterogeneous graph embedding method, The choice of meta path is still an open and challenging problem. the design of the meta-path scheme largely depends on domain knowledge. in this paper, We propose a heterogeneous graph neural network with denoising (HGNND) to handle the issue. considering that there are different types of nodes in heterogeneous graphs, And their features are usually distributed in different spaces, The HGNND projects features of different types of nodes into a common vector space. then, The whole heterogeneous graph is input into the graph neural network to aggregate the neighbor node information and capture the structure information of the heterogeneous graph. finally, The noise nodes that may affect the performance of the whole model are filtered out by the denoising operation. extensive experiments on three real-world datasets demonstrate that our proposed model achieves state-of-the-art performance, It further proves that the model can still effectively aggregate semantic information without using meta-paths. (c) 2021 Elsevier B.V. All rights reserved.

Automated summary

In this paper, a heterogeneous graph neural network with denoising (HGNND) is proposed to address the issue of meta-path selection in graph embedding. By projecting the features of different types of nodes into a common vector space and utilizing a graph neural network to aggregate neighbor node information and capture the structure information of the heterogeneous graph, the proposed model achieves state-of-the-art performance on three real-world datasets.