Link communities detection: an embedding method on the line hypergraph

Abstract Recent advances have verified ground-truth communities perceive several characteristics. That is, communities are overlapped and densely connected. Not only that, the organization of communities, in a general sense, is hierarchical. To capture all of these characteristics, we propose a framework based on link embedding method. Firstly, we define close-knit link groups which preserve the hierarchical structures and carefully transform the problem of mining close-knit link groups as mining cosine patterns which can be implemented efficiently. Secondly, we construct the weighted line hypergraph and embed each link into a low dimension vector. Finally, we simply employ K-means algorithm to obtain the link communities. Overlapping structures are naturally obtained by interpreting the link communities as nodes communities. Experimental results on three real-world networks demonstrate the proposed approach is able to identify much higher-quality overlapping communities in terms of four external measures, compared with six classical overlapping community detection methods.