Link Constraint Reliability Learning for Semi-Supervised Community Detection

Community detection is vital to understand structures and functions in complex networks. Current community detection methods by network topology alone show their inability to handle noise in networks, and these methods are problematic for applications that impose constraints on the acceptable solutions. To address these problems, we develop a new technique by utilizing prior information and network links, called link constraint reliability learning, for community detection. Unlike most existing semi-supervised methods, only considering the given constraints and ignoring their impacts on other underlying constraints, the core idea of our approach is to infer new constraints from the links in original networks by assessing their reliability level. Specifically, we first select the must-link priors as the seeds and rank all the links using manifold ranking model. And then the links with low reliability level are inferred as new cannot-link constraints. Finally both the given constraints and derived constraints are encoded into the adjacency matrix. The use of new constraints inference significantly de-noise the consensus matrix of community structures, thus the prior information is more fully exploited. Extensive experiments on real-world networks demonstrate that the proposed method outperforms state-of-the-art approaches, especially on networks with unclear structures.