Supervised Link Prediction Developed For Bipartite Social Networks

Abstract

Link prediction is a prominent issue that involves predicting the occurrence of future relationships between nodes in a social network. Our work proposes similarity metrics that are extended for a weighted bipartite social network to predict prospective links in the social network by applying a supervised machine learning scheme. Link (edge) weights in the network could provide valuable information for prediction as they express the strength of relationships between nodes (person, item etc.). The target attribute of prediction is a label that shows the existence or absence of a link between two nodes in the network. The feature attributes of the machine learning model are similarity/centrality metrics calculated from the current social network. Particularly, a weighted bipartite graph was built from the MovieLens dataset by connecting users to movies via the users' movie ratings; then new links were attempted to predict for a later time. Several types of machine learning algorithms for link prediction on this bipartite graph were applied by using network similarity metrics and a binary supervised classifier. The combination of four network centrality metrics provided higher prediction performance compared their individual performances on the bipartite movie ratings network. Our preliminary experiments led satisfactory results when link weights were considered, which encourages us for further analysis on bipartite and weighted social networks.