Model-Based Clustering for Large-Scale Dynamic Networks

• Activity Number: 614
• Type: Contributed
• Date/Time: Wednesday, August 12, 2015 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Kevin Lee* and Lingzhou Xue and David R. Hunter
• Companies: Penn State and Penn State and Penn State
• Keywords: Community detection ; Large-scale dynamic network ; EM algorithm

Abstract

Community detection is a fundamental research topic for exploring large-scale networks. In real-world dynamic networks, it is important to identify overlapping communities because nodes are naturally characterized by multiple community memberships. In this work, we propose a unified model-based clustering scheme for large-scale dynamic exponential-family random graph models. In particular, we employ the parsimony and flexibility of the conditional dyadic independencies given unobserved memberships to address the scalability of exponential-family distributions for modeling large dynamic networks. The proposed method effectively detects overlapping communities for large-scale temporal exponential-family random graph models (Hanneke et al. 2010) and separable temporal exponential-family random graph models (Krivitsky & Handcock, 2014). Moreover, we design an efficient variational generalized EM algorithm to implement the proposed method. The numerical performance of the proposed method is demonstrated in both simulation studies and real applications.