Efficiently and Fast Learning a Fine-grained Stochastic Blockmodel from Large Networks

Abstract

Stochastic blockmodel (SBM) has recently come into the spotlight in the domains of social network analysis and statistical machine learning, as it enables us to decompose and then analyze an exploratory network without knowing any priori information about its intrinsic structure. However, the prohibitive computational cost limits SBM learning algorithm with the capability of model selection to small network with hundreds of nodes. This paper presents a fine-gained SBM and its fast learning algorithm, named FSL, which ingeniously combines the component-wise EM (CEM) algorithm and minimum message length (MML) together to achieve the parallel learning of parameter estimation and model evaluation. The FSL significantly reduces the time complexity of the learning algorithm, and scales to network with thousands of nodes. The experimental results indicate that the FSL can achieve the best tradeoff between effectiveness and efficiency through greatly reducing learning time while preserving competitive learning accuracy. Moreover, it is noteworthy that our proposed method shows its excellent generalization ability through the application of link prediction.