A0551
Title: Dimension reduction and changepoint detection in network series
Authors: Michael Weylandt - University of Florida (United States) [presenting]
Abstract: As social networks and Internet of Things systems become increasingly common, the analysis of network data, and data observed on those networks, holds great potential but poses several acute statistical challenges. Foremost among these is the small sample sizes typically associated with network data, often far less than the large scale and high-dimensionality of the systems of interest. To address this, we propose a framework for dimension reduction of networks observed over time and apply it to changepoint detection. The framework is flexible, allowing for both parametric and non-parametric models of the underlying network dynamics to be used. We prove the consistency of the proposed approach under several popular network models and provide efficient tensor decomposition algorithms suitable for use on large-scale networks. As a byproduct of our analysis, we present several new consistency results for high-dimensional tensor decompositions, which are likely to be of independent interest.
