HGNAC-HS: Heterogeneous graph node attribute completion method integrating high-order semantics

In recent years, heterogeneous graphs (HGs) have gradually attracted research interest due to their ability to effectively describe and model complex relational networks in real-world. Aiming at this complex graph structure, researchers have proposed a series of excellent heterogeneous graph neural networks (HGNNs), however, many real-world scenarios involving HGs usually suffer from the problem of missing node attributes. In addition, existing HGNNs with attribute complementation often require pre-training to acquire a priori knowledge before performing attribute complementation, and the higher-order semantic information is underutilized. To this end, we propose an attribute complementation framework (HGNAC-HS). Specifically, we first design an encoder to learn complex graph structures and higher-order semantic information; then, based on capturing the higher-order semantics, we use the attention mechanism to complement the missing attributes of the nodes; and finally, we combine HGNAC-HS with the MAGNN and HGT models to make the whole model end-to-end. In order to verify the effectiveness of our work, we conduct a large number of experiments on ACM and IMDB datasets, which demonstrate that the proposed HGNAC-HS exhibits superior performance.