ReAHGN: Adaptive Heterogeneous Graph Neural Network With Relation-Aware Embedding

Abstract

Heterogeneous graph neural networks (HGNs) have attracted more and more attention recently due to their wide applications such as node classification, community detection, and recommendation. Significant progress has been made by graph neural networks. However, we argue that attention mechanisms that focus on computing higher ranking scores for specific types of nodes cannot select the most relevant neighbors for any target node. To enhance the expressiveness of the HGNs, we propose ReAHGN, a Relation-aware embedding for Adaptive Heterogeneous Graph Neural Network. Specifically, we present an adaptive attention mechanism that assigns different weights for each type of target node automatically. In addition, a relation-aware embedding method is adopted to fuse the edge type information effectively. ReAHGN is evaluated for three downstream tasks and on ten public datasets. Experiment results demonstrate that ReAHGN can outperform the state-of-the-art HGNs.