Meta-path-based link prediction in schema-rich heterogeneous information network

Abstract

Recent years have witnessed the boom of heterogeneous information network (HIN), which contains different types of nodes and relations. Complex structure and rich semantics are unique features of HIN. Meta-path, the sequence of object types and relations connecting them, has been widely used to mine this semantic information in HIN. Link prediction is an important data mining task to predict the potential links among nodes, which are required in many applications, e.g., filling missing links. The contemporary link prediction is usually based on simple HIN whose schema is bipartite or star schema. In these works, the meta-paths should be predefined or enumerated. However, in many real networked data, it is hard to describe their network structures with simple schema. For example, the RDF-formatted Knowledge Graph which includes tens of thousands types of objects and links is a kind of schema-rich HIN. In this kind of schema-rich HIN, it is impossible to enumerate meta-paths so that the contemporary work is invalid. In this paper, we study link prediction in schema-rich HIN and propose a novel method named Link Prediction with automatic meta Path (LiPaP). The LiPaP designs an algorithm called automatic meta-path generation to automatically extract meta-paths from schema-rich HIN in the approximate order of relevance and adopt a supervised method with likelihood function to learn the weights of extracted meta-paths. Extensive experiments on real knowledge database, Yago, demonstrate that LiPaP is an effective, steady and efficient approach.