AHNG: Representation learning on attributed heterogeneous network

Network embedding aims to encode nodes into a low-dimensional space with the structure and inherent properties of the networks preserved. It is an upstream technique for network analyses such as link prediction and node clustering. Most existing efforts are devoted to homogeneous or heterogeneous plain networks. However, networks in real-world scenarios are usually heterogeneous and not plain, i.e., they contain multi-type nodes/links and diverse node attributes. We refer such kind of networks with both heterogeneities and attributes as attributed heterogeneous networks (AHNs). Embedding AHNs faces two challenges: (1) how to fuse heterogeneous information sources including network structures, semantic information and node attributes; (2) how to capture uncertainty of node embeddings caused by diverse attributes. To tackle these challenges, we propose a unified embedding model which represents each node in an AHN with a Gaussian distribution (AHNG). AHNG fuses multi-type nodes/links and diverse attributes through a two-layer neural network and captures the uncertainty by embedding nodes as Gaussian distributions. Furthermore, the incorporation of node attributes makes AHNG inductive, embedding previously unseen nodes or isolated nodes without additional training. Extensive experiments on a large real-world dataset validate the effectiveness and efficiency of the proposed model.