Link prediction in weighted networks via motif predictor

Abstract

Because of its very importance in a plethora of applications, the research of link prediction has received growing attention from different disciplines. However, link prediction in weighted networks is still a challenge owning to the weak tie phenomenon. To address the issue, this paper proposes a new method for link prediction, which can adaptively assess the connection likelihood of node pairs. In the proposed method, all links are classified into four types according to their weights, and ten pairs of triadic motifs are attained based on this partition. Then, a pair of triadic motifs are considered as a motif predictor to gauge the connection probability of two unconnected nodes using the naïve Bayes model. At last, the final connection likelihood score of these two nodes is computed by summing the probabilities got from all predictors. The performance of the proposed method is studied by performing several experiments on a number of real-world and synthetic weighted networks. The results show that the accuracy of the proposed method outperforms that of the compared methods in most cases.