Continuous-Time Relationship Prediction in Dynamic Heterogeneous Information Networks

摘要

Online social networks, World Wide Web, media and technological networks, andother types of so-called information networks are ubiquitous nowadays. Theseinformation networks are inherently heterogeneous and dynamic. They areheterogeneous as they consist of multi-typed objects and relations, and theyare dynamic as they are constantly evolving over time. One of the challengingissues in such heterogeneous and dynamic environments is to forecast thoserelationships in the network that will appear in the future. In this paper, wetry to solve the problem of continuous-time relationship prediction in dynamicand heterogeneous information networks. This implies predicting the time ittakes for a relationship to appear in the future, given its features that havebeen extracted by considering both the heterogeneity and the temporal dynamicsof the underlying network. To this end, we first introduce a meta-path-basedfeature extraction framework to effectively extract features suitable forrelationship prediction regarding the heterogeneity and dynamicity of thenetwork. Next, we propose a supervised nonparametric approach, calledNon-Parametric Generalized Linear Model (NP-GLM), which infers the hiddenunderlying probability distribution of the relationship building time given itsfeatures. We then present a learning algorithm to train NP-GLM and an inferencemethod to answer time-related queries. Extensive experiments conducted on bothsynthetic dataset and real-world DBLP bibliographic citation network datasetdemonstrate the effectiveness of Np-Glm in solving continuous-time relationshipprediction problem vis-a-vis alternative baselines.