Using clustering coefficient to construct weighted networks for supervised link prediction

Abstract

Link prediction problem has drawn much attention in the analysis of complex networks. A lot of previous works are devoted for performing link prediction both on weighted and un-weighted networks. Clustering coefficient is a well-studied attribute in graph theory. It measures the degree to which nodes in a graph tend to cluster together. In this paper, we apply normalized clustering coefficient as a weighting scheme to construct weighted networks for supervised link prediction. Unlike the weighting approaches used in some existing literatures which have to consider properties outside the scope of network topology, we take only the topological properties into consideration. Several proximity metrics like Common Neighbors, Jaccard’s coefficient and Adamic–Adar coefficient are chosen as predictor attributes. Learning algorithms like Naive Bayes, J48 and SVM available in the WEKA environment are used to perform binary classification experiments on the constructed networks. Both weighted and un-weighted versions of networks are used and experimental results show that the proposed weighting approach could bring benefits to link prediction.