Label Preserved Heterogeneous Network Embedding

Recently, the heterogeneous network embedding (HNE for short) methods have been attracting increasing attention due to their simplicity, scalability, and effectiveness. However, the rich node label information is not considered by these HNE methods, which leads to suboptimal node embeddings. In this paper, we propose a novel Label Preserved Heterogeneous Network Embedding (LPHNE) method to tackle this problem. Briefly, for each type of the nodes, LPHNE projects these nodes and their labels into a same low-dimensional hidden space by modeling the interactive relationship between the labels and the contexts of the nodes. Thus, the discriminability of node embedding is improved by utilizing the label information. The extensive experimental results demonstrate that our semi-supervised method outperforms the various competitive baselines on two widely used network datasets significantly.