Finding missing links in an observed network (static) or predicting those links that may appear in the future (dynamic) is the aim of the link prediction (LP) task. LP plays a significant role in network evolution, as shown by several works (Barabasi and Albert,1999; Kleinberg, 2000; Leskovec et al., 2005). However, this problem is still challenging for the authors. Most approaches are based on the topological properties of networks like degree, clustering coefficient, path index, etc. The common neighbor approaches are based on the idea "Friends of a friend are also friends,'' i.e., a large number of common friends between two persons (nodes) signifies more similarity between them and more likely to be connected. In the resource allocation process, a large number of connections of common neighbors of two nodes are vulnerable for leaking information (resources) through them. Based on this idea, we proposed a new similarity index SHOPI (Link prediction based on Significance of Higher Order Path Index) that tries to constrain the information leak through the common neighbors by penalizing them. Moreover, higher-order paths (as defined by six degrees of separation) are used as discriminating features with penalizing the longer paths available between the seed node pair. The experimental results on twelve real-world network datasets (collected from diverse areas) show that SHOPI outperforms the baseline methods. Moreover, SHOPI is more robust than the existing Katz index and comparable to the local path index (LP). The statistical test shows the significant difference of the proposed method (i.e., SHOPI) with the baseline approaches. (C) 2019 Elsevier B.V. All rights reserved.
