A Comparative Study of Link Prediction Algorithms For Social Networks of Varying Sizes

Abstract

Link prediction in social networks enables us to predict future links in an evolving social network as new nodes get added through the passage of time. It can also help to detect missing edges in the social graph. A successful link prediction method substantially reduces the experimental effort required to establish the topology of a network and can accelerate the mutually beneficial interactions that takes much longer to form by chance. In this work, the various state of art link prediction methods are analyzed while also looking at the kind of networks the algorithms work best with. Different social graph data sets are chosen based on their sizes and sparsity. Subsequently link prediction task on each of these data sets were done using various machine learning algorithms to understand the performance of several random walk based node embedding methods as compared with the performance of several classical approaches. It is found that the sparsity and size of the graph are important factors that determine the performance of random walk based node embedding methods.