An End-To-End Heterogeneous Graph Clustering Method

Heterogeneous graph clustering is a fundamental and difficult task in data mining. It is a great challenge to complete the clustering process while preserving the structural information of heterogeneous graphs. Therefore, an end-to-end heterogeneous graph clustering method is proposed, which aims to jointly and optimally learn the heterogeneous graph node representation process and the clustering process. Specifically, we use heterogeneous graph auto-encoders to model heterogeneous graphs and learn their node representations. Meanwhile we jointly guiding the generation of node representations by constructing auxiliary distributions oriented to clustering. So, the learned node representations not only preserve the structural information of heterogeneous graphs, but also make them separate in the vector space for clustering purposes. The experimental results show that the joint learning of node representations and clustering has better performance than the traditional separate learning method.