Efficiently detecting overlapping communities using seeding and semi-supervised learning

A common scheme for discovering overlapping communities in a network is to use a seeding process followed by an expansion process. Most seeding methods are either too complex to scale to large networks or too simple to select high-quality seeds. Additionally, the non-principled functions used by most expansion methods lead to poor performances when applied to diverse networks. This paper proposes a new method that transforms a network into a corpus. Each edge is treated as a document, and all the network nodes are treated as terms of the corpus. We propose an effective seeding method that selects seeds as a training set, and a principled expansion method based on semi-supervised learning that classifies the edges. We compared our new algorithm with four other community detection algorithms on a wide range of synthetic and empirical networks. Our experimental results show that the new algorithm significantly improved the clustering performance in most cases. Furthermore, the time complexity of the new algorithm is linear with respect to the number of edges, which means that the technique can be scaled to large networks.