Simple and Efficient Heterogeneous Graph Neural Network

Heterogeneous graph neural networks (HGNNs) have the powerful capability to embed rich structural and semantic information of a heterogeneous into node representations. Existing HGNNs inherit many mechanisms from (GNNs) designed for homogeneous graphs, especially attention mechanism multi-layer structure. These bring excessive complexity, but seldom work studies whether they are really effective on graphs. In this paper, we conduct an in-depth detailed study these propose Simple Efficient Graph Neural Network (SeHGNN). To easily capture information, SeHGNN pre-computes neighbor aggregation using light-weight mean aggregator, which reduces complexity by removing overused avoiding repeated in every training epoch. better utilize adopts single-layer structure with long metapaths extend receptive field, as well transformer-based fusion module fuse features different metapaths. As result, exhibits characteristics simple network structure, high prediction accuracy, fast speed. Extensive experiments five real-world graphs demonstrate superiority over state-of-the-arts both accuracy