Similarity Measures for Link Prediction Using Power Law Degree Distribution

Abstract

Link Prediction deals with predicting important non-existent edges in a social network that are likely to occur in the near future. Typically a link between two nodes is predicted if the two nodes have high similarity. Well-known local similarity functions make use of information from all the local neighbors. These functions use the same form irrespective of the degree of the common neighbors. Based on the power law degree distributions that social networks generally follow, we propose non-linear schemes based on monotonically non-increasing functions that give more emphasis to low-degree common nodes than high-degree common nodes. We conducted experiments on several benchmark datasets and observed that the proposed schemes outperform the popular similarity function based methods in terms of accuracy.