Self-Supervised Heterogeneous Graph Neural Network Based on Multi-Head Self-Attention

In this study, a self-supervised heterogeneous graph neural network called MultiHeCo is proposed, which introduces a multi-head self-attention mechanism to enhance the ability of heterogeneous graph neural networks to process complex interactions. For the complex structure of heterogeneous information networks, the multi-head self-attention mechanism is able to process and capture complex information in parallel. Since this mechanism is suitable for capturing the characteristics of long-distance dependencies, this study applies it to meta-path view generation to identify key features and relationships. Specifically, we introduce the multi-head self-attention mechanism in heterogeneous graph neural networks for the first time, propose a new meta-path view node embedding structure, and validate it on two public datasets. The experimental results show that the model achieves the highest level in several downstream tasks, even surpassing the performance of some semi-supervised learning models.