A Comparative Study of Classification Algorithms for Link Prediction

A graph structure represents various networks; such as social networks, bibliographic networks, protein-protein interactions and communication networks. Determining possible links in a network is one of the most prevalent tasks which is defined as the link prediction problem. Finding probable links for data modeled as networks is used for predictive systems and recommendation systems. In this paper, we compare different supervised and unsupervised classification algorithms on their performance to accurately predict missing links. First, we model our network by creating baseline parameters such as adamic adar, resource allocation index, preferential attachment index and common neighbors. The approach is tested on Facebook dataset for one ego-net network. The performance of every model is evaluated on its area under curve, receiver operating characteristic score where we observe that support vector machine performs better than all the other supervised learning algorithms.