Automated Bisimulation-Based Similarity Measurement in Heterogeneous Information Networks

Abstract

Heterogeneous information networks (HINs) serve as effective models for information systems characterized by diverse types of objects and relationships. Evaluating similarities among objects is crucial in various data mining applications, such as web search, label prediction, and clustering tasks. This paper presents BiSim, a novel similarity measurement method tailored for HINs. By harnessing the concept of bisimulation, BiSim evaluates node similarity by integrating both macroscopic and microscopic levels of bisimulation. Unlike existing metrics that rely on predefined metapaths, BiSim provides a universal approach to assess the structural and semantic similarity simultaneously in HINs. We thoroughly investigate BiSim's mathematical properties and demonstrate its effectiveness through comprehensive experimentation across diverse data mining tasks.