We tackle the task of link prediction in social networks as a supervised learning problem. We construct nodebased, neighborhood-based, path-based and community detection features, as well as a more recent node2vec approach that is based on random walks. We feed these features into multiple machine learning algorithms and evaluate the results on both directed and undirected social networks. We additionally attempt to explain the fitted models and analyze whether random sampling of edges can really give us a good model for predicting future edges, which is usually the desired task for the purpose of friends recommendation systems.
Paper: https://github.com/simejanko/link-prediction/blob/master/paper.pdf