Relation Heterogeneous Graph Neural Network

Abstract

In heterogeneous graph, we can mine high-order neighbor information or semantic information using meta-path, or only use the original connection, and then obtain high-order neighbor information indirectly through residual connections. Both two methods can get good results, but the latter can improve the efficiency without prior knowledge and meta-path mining. We take the second approach, proposing a novel relation heterogeneous graph neural network (RHGN) which adds edge features to the message aggregation of nodes and updates edge information by comparing different edge types through auxiliary tasks. Extensive experiments on two real-world heterogeneous graphs of node classification tasks show that our proposed model works better.