Utilizing adjacency of colleagues and type correlations for enhanced link prediction

Abstract Discoveries of new relationships in the network of objects have been required in various applications such as social networks, DBLP bibliographic networks and biological networks. Specifically, link prediction in heterogeneous information networks (HINs) that consist of multiple types of nodes and links has received much attention recently because many information networks of the real world are HINs. We observe various factors that affect the existence of a link in HINs. Firstly, certain structural characteristics of nodes whose types are the same as that of a source (or target) node give important information for link prediction. Secondly, in the HINs, there can be meaningful correlation between links of a particular link type and paths of a particular path type (also called a meta-path). In other words, paths of different path types affect the existence of links differently. Finally, we use the number of paths between source and target nodes to measure proximity of two nodes. Based on these observations, we newly propose several features and a prediction model. We show through various experiments that our proposed method works effectively and performs better than the other existing methods.