Newton’s gravitational law for link prediction in social networks

Authors: Wahid-Ul-Ashraf, A., Budka, M. and Musial-Gabrys, K.

http://eprints.bournemouth.ac.uk/30084/

Start date: 29 November 2017

This data was imported from DBLP:

Authors: Wahid-Ul-Ashraf, A., Budka, M. and Musial-Gabrys, K.

Editors: Cherifi, C., Cherifi, H., Karsai, M. and Musolesi, M.

http://eprints.bournemouth.ac.uk/30084/

https://doi.org/10.1007/978-3-319-72150-7

Journal: COMPLEX NETWORKS

Volume: 689

Pages: 93-104

Publisher: Springer

ISBN: 978-3-319-72149-1

This data was imported from Scopus:

Authors: Wahid-Ul-Ashraf, A., Budka, M. and Musial-Gabrys, K.

http://eprints.bournemouth.ac.uk/30084/

Journal: Studies in Computational Intelligence

Volume: 689

Pages: 93-104

ISBN: 9783319721491

ISSN: 1860-949X

DOI: 10.1007/978-3-319-72150-7_8

© Springer International Publishing AG 2018. Link prediction is an important research area in network science due to a wide range of real-world application. There are a number of link prediction methods. In the area of social networks, these methods are mostly inspired by social theory, such as having more mutual friends between two people in a social network platform entails higher probability of those two people becoming friends in the future. In this paper we take our inspiration from a different area, which is Newton’s law of universal gravitation. Although this law deals with physical bodies, based on our intuition and empirical results we found that this could also work in networks, and especially in social networks. In order to apply this law, we had to endow nodes with the notion of mass and distance. While node importance could be considered as mass, the shortest path, path count, or inverse similarity (AdamicAdar, Katz score etc.) could be considered as distance. In our analysis, we have primarily used degree centrality to denote the mass of the nodes, while the lengths of shortest paths between them have been used as distances. In this study we compare the proposed link prediction approach to 7 other methods on 4 datasets from various domains. To this end, we use the ROC curves and the AUC measure to compare the methods. As the results show that our approach outperforms the other 7 methods on 2 out of the 4 datasets, we also discuss the potential reasons of the observed behaviour.

The data on this page was last updated at 04:59 on September 22, 2018.