Link Prediction in Temporal Heterogeneous Networks

Abstract

Link prediction in temporal social networks addresses the problem of predicting future links. The problem of link prediction in heterogeneous networks is challenging due to the existence of multiple types of nodes and edges. There are many methods available in the literature for homogeneous networks, which rely on the network topology. In this work, we extend some of the standard measures viz Common Neighbors, Jaccard Coefficient, AdamicAdar, Time-score, Co-occurrence probabilistic measure and Temporal Co-occurrence probabilistic measure to heterogeneous networks. Probabilistic graphical models prove to be efficient for link prediction compared to topological methods. We incorporate the information related to time of link formation into probabilistic graphical models and generate a new measure called Heterogeneous Temporal Co-occurrence probability (Hetero-TCOP) measure for heterogeneous networks. We evaluate all the extended heterogeneous measures along with Hetero-TCOP on DBLP and HiePh bibliographic networks for predicting two types of links: author-conference/journal links and co-author links in the heterogeneous environment. In both cases, Hetero-TCOP achieves superior performance over the standard topological measures. In the case of DBLP dataset, Hetero-TCOP shows an improvement of 15% accuracy over neighborhood-based measures, 6% over temporal measures and 5% over Co-occurrence probability measure. Similar improvement in performance is observed for HeiPh dataset also.