A new study of using temporality and weights to improve similarity measures for link prediction of social networks

Article type: Research Article

Authors: Aghabozorgi, Farshad | Reza Khayyambashi, Mohammad^{; *}

Affiliations: Faculty of Computer Engineering, University of Isfahan, Isfahan, Iran

Correspondence: [*] Corresponding author. Mohammad Reza Khayyambashi, Faculty of Computer Engineering, University of Isfahan, Postal Code 81746-73441, Isfahan, Iran. Tel.: +98 913 3676728; E-mail: m.r.khayyambashi@comp.ui.ac.ir.

Abstract: Link prediction is the problem of inferring future interactions among existing network members based on available knowledge. Computing similarity between a node pair is a known solution for link prediction. This article proposes some new similarity measures. Some of them use nodes’ recency of activities, some weights of edges and some fusion of both in their calculation. A new definition of recency is provided here. A supervised learning method that applies a range of network properties and nodes similarity measures as its features set is developed here for experiments. The results of the experiments indicate that using proposed similarity measures would improve the performance of the link prediction.

Keywords: Link prediction, supervised learning, recency, similarity measures, social networks

DOI: 10.3233/JIFS-17770

Journal: Journal of Intelligent & Fuzzy Systems, vol. 34, no. 4, pp. 2667-2678, 2018