Heterogeneous graph embedding by aggregating meta-path and meta-structure through attention mechanism

Heterogeneous information networks embedding, which is a promising technique to learn low-dimensional representations for nodes with different types, has obtained very high interests recently. Plenty of graph neural networks models have been proposed for heterogeneous graph embedding. However, there are two limitations in existing models: (1) despite the complex structures of heterogeneous nodes, almost all these models are mainly based on meta-paths; (2) the different importance of different structures are neglected in most models. To address these problems, we propose a meta-path and meta-structure integrated heterogeneous graph neural network through attention mechanisms (PSHGAN). PSHGAN first maps features of heterogeneous nodes into the same space. Then, PSHGAN learns the weights of two nodes at each end of the meta-path or meta-structure by a local attention mechanism. Finally, PSHGAN learns weights of meta-paths and meta-structures by a global attention mechanism and aggregates the nodes representations. Extensive experiments are conducted on real-world benchmark datasets and show that our proposed model outperforms the state-of-the-art models in the node classification and link prediction tasks. Moreover, we make comprehensive analysis on the impacts of meta-structures on the performance of classification training with meta-paths.