Heterogeneous Graph Neural Networks Based on Meta-path

Heterogeneous graph representation learning aims to learn meaningful representation vectors from heterogeneous networks in low dimension, so as to realize the extraction of structure and attribute features of the networks. Embedding vector is the basis and key of complex network analysis, which can be used in the downstream tasks. The key points in heterogeneous graph neural networks are: how to define heterogeneous neighbors and how to aggregate them. Although a lot of work has been devoted to homogeneous or heterogeneous network representation, the effective combination of network structure information and node attribute information, especially the effective use of meta-path containing specific semantic information is still rare. In this paper, we propose a meta-path-based heterogeneous graph neural network model. Firstly, we apply meta-path to sample the heterogeneous neighbors of each node in the network, and aggregate the features of the same type of nodes together to form type-related embedding; then, attention mechanism is applied to aggregate the neighbor information of different types of node; finally we train the end-to-end model by reducing the context loss. Experiments proved the validity of the model and significantly improved current results.