Robust stochastic block model

Abstract The family of stochastic block model (SBM) is a mainstay to detect network structures, especially for the exploratory networks analysis without any prior. However, the real-world networks often contain many noisy nodes that have abnormal behaviors or go against the certain patterns. This creates the so-called noise problem, resulting in lower performance of SBMs in real applications. To alleviate this problem, we propose a novel Robust Stochastic Block Model (RSBM). The proposed method can model the noisy nodes in the network and maintain the ability of SBM in structure analysis. RSBM is inferred using variational Bayesian expectation maximization. We evaluate RSBM on both synthetic and real-world networks, and empirical results demonstrate that our RSBM outperforms the state-of-the-art baseline models in the structural partitioning task.