Title: Modeling and Detection using High-dimensional Time Series Data

Date: May 1st (Thursday)

Time: 2 pm – 4 pm EST

Location: ISyE Studio 109

Zoom link: https://gatech.zoom.us/j/95226220687 

Tingnan Gong

Industrial Engineering PhD Student

H. Milton Stewart School of Industrial and Systems Engineering (ISyE)
Georgia Institute of Technology

Committee

1 Dr. Yao Xie (Advisor, ISyE)

2 Dr. Seong-Hee Kim (Advisor, ISyE)

3 Dr. Kamran Paynabar (ISyE)

4 Dr. Nagi Gebraeel (ISyE)

5 Dr. Simon Mak (Duke, Statistical Science)

Abstract

Change-point detection and point process modeling play a vital role in time series analysis, particularly when the data are high-dimensional, correlated, or observed under time uncertainty. For high-dimensional data with complex spatio-temporal correlations and minimal distributional assumptions, real-time detection of distributional shifts requires carefully designed procedures that address multiple statistical and computational challenges. For event data with time uncertainty, to the best of our knowledge, no existing work has established a point process model that explicitly incorporates time uncertainty. This thesis develops new methods for online change-point detection in complex high-dimensional data and for point process modeling under time uncertainty.

Chapter 1 proposes a distribution-free CUSUM procedure for low-rank image data. The method avoids parametric noise assumptions and demonstrates reliable detection performance in manufacturing applications. Chapter 2 introduces a change-point detection procedure that trains a neural network-based statistic via a binary classification proxy. This neural CUSUM detector adapts flexibly to various types of distributional shifts. Chapter 3 addresses the detection of sparse and weak changes in high-dimensional signals using a higher-criticism-based approach. The proposed method achieves asymptotically optimal detection delay under rare moderate departure regimes. Chapter 4 develops a novel framework for modeling point processes with time uncertainty. Through a carefully designed training scheme and kernel-based parameterization, the model can predict event occurrence probabilities and recover dynamic causal structures in both simulations and real-world datasets such as Sepsis medical records and urban burglary incidents.