Learning Network Dynamics via Regularized Tensor Decomposition

• Activity Number: 131
• Type: Contributed
• Date/Time: Monday, August 1, 2016 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistical Learning and Data Science
• Author(s): Yun-Jhong Wu* and Elizaveta Levina and Ji Zhu
• Companies: University of Michigan and University of Michigan and University of Michigan
• Keywords: dynamic networks ; non-negative tensor decomposition ; under-complete tensor representation ; power method

Abstract

Real networks often evolve over time, and interactions between nodes in networks are usually observed only at certain specific time points. In this work, we consider network data with time-stamped links. We propose to model such a dynamic network using a low rank tensor representation. This model characterizes time trends of multiple rank-1 factors and can be used to approximate more complicate networks. We develop an approach to fit this model based on a tensor completion algorithm and a smoothness penalty in the time domain, implemented with a highly scalable power-iteration-based algorithm which can fit large sparse dynamic networks. The numerical experiments on simulated data as well as the Enron e-mail dataset demonstrate the potential of tensor methods for dynamic network data analysis.