Title

Data Driven Network Segmentation and Compression

Time

1:40pm - 3:00pm

Location

1605 Tilia, Room 1103, West Village

Abstract

Underdetermined systems are commonplace in transportation network analysis, as models become increasingly non-parametric. Variable fusion provides a solution to this problem by adding penalty terms to optimization-based statistical methods that fuse parameters in a data driven fashion. Total variation denoising, also known as the fused lasso, is one such tool that is employed in computer vision for segmenting images. In the first project, we adapt these methods to segment networks for density estimation over roads. We will discuss recent advances in optimizing the resultant objectives, and outline potential applications. In the second project, we will discuss spatio-temporal network data and use total variation denoising tools for segmenting both space and time for better demand estimation and dataset compression. These methodologies have applications to auto insurance rating, ride-sharing services, and designing transportation policies.

Biographical Sketch

James is an Assistant Professor of Statistics at UC Davis.  He received his Ph.D. in Machine Learning and Statistics from Carnegie Mellon University.  He designs and studies efficient methods for statistical problems involving networks, which includes regression on a graph, estimating network models, and unsupervised methods for visualizing and exploring networks.  He is a father, avid hiker, and loves history.