Key Nodes Cluster Augmented Embedding for Heterogeneous Information Networks

Heterogeneous Information Networks (HINs), composed of multiple types of node and relation, usually have more expressive ability for complex relational data. Recently, network embedding aiming to project the network into a low-dimensional vector space has received much attention. Most existing embedding methods for HINs utilize meta-path to capture the proximity of node. However, these methods usually ignore the inequivalence of different types of node and clustering structure of network, which are important characteristics of HINs. Hence, we propose a key node based heterogeneous network embedding method enhanced by the clustering information. In our method, we first utilize a meta-path guided random walk to obtain general node representations in terms of rich heterogeneous semantic features in HINs. To indicate different equivalence of nodes, we define key nodes which are usually in the essential location in HINs, such as paper in the bibliographic HINs. Afterwards, we incorporate the clustering structure of the key nodes into network embedding learning via Gaussian Mixture Model to further enhance the representations of nodes. Lastly, we design a unified objective function to mutually learn the two parts effectively. Extensive experiments are conducted and the results validate the effectiveness of our model.