Online Program Home
  My Program

All Times EDT

Abstract Details

Activity Number: 583 - Learning Network Structure in Heterogeneous Populations
Type: Topic Contributed
Date/Time: Thursday, August 6, 2020 : 3:00 PM to 4:50 PM
Sponsor: Section on Statistical Learning and Data Science
Abstract #313404
Title: Optimistic Binary Segmentation with an Application in Change Point Detection Methodologies for Graphical Models in the Presence of Missing Values
Author(s): Solt Kovács* and Peter Bühlmann and Lorenz Haubner and Housen Li and Malte Londschien and Axel Munk
Companies: ETH Zurich and ETH Zurich and ETH Zurich and University of Göttingen and ETH Zurich and University of Göttingen
Keywords: Fast computation; Incomplete and heterogeneous data; Time-varying models; Break points; High-dimensional; Precision and covariance matrix estimation

High-dimensional change point detection problems pose several challenges. First, we discuss computational challenges whenever underlying model fits are expensive and propose a widely applicable new methodology, called optimistic binary segmentation, that has the potential for massive computational savings while maintaining good statistical performance. In the second part, we discuss the commonly occurring phenomenon of missing values and why this poses particular challenges in heterogeneous populations, i.e., in the presence of change points. Specifically, we propose estimation methods for change points in high-dimensional covariance structures and advocate three imputation methods for scenarios with missing values. We investigate implications on common losses used for change point detection, with an emphasis on the applicability of optimistic binary segmentation in such challenging scenarios with missing values. We also discuss how model selection methods have to be adapted to the setting of incomplete data. The methods are compared in a simulation study and applied to detect changes in time series from an environmental monitoring system with naturally occurring missing values.

Authors who are presenting talks have a * after their name.

Back to the full JSM 2020 program