Fine-Grained Semantics-Aware Heterogeneous Graph Neural Networks

Designing a graph neural network for heterogeneous graph which contains different types of nodes and links have attracted increasing attention in recent years. Most existing methods leverage meta-paths to capture the rich semantics in heterogeneous graph. However, in some applications, meta-path fails to capture more subtle semantic differences among different pairs of nodes connected by the same meta-path. In this paper, we propose Fine-grained Semantics-aware Graph Neural Networks (FS-GNN) to learn the node representations by preserving both meta-path level and fine-grained semantics in heterogeneous graph. Specifically, we first use multi-layer graph convolutional networks to capture meta-path level semantics via convolution on edge type-specific weighted adjacent matrices. Then we use the learned meta-path level semantics-aware node representations as guidance to capture the fine-grained semantics via the coarse-to-fine grained attention mechanism. Experimental results semi-supervised node classification show that FS-GNN achieves state-of-the-art performance.