Doctoral Dissertation Proposal Oral Defense, Huiqun Huang

Title: Machine Learning and Uncertainty Quantification for Spatiotemporal Data Prediction
Ph.D. Candidate: Huiqun Huang
Major Advisor: Dr. Fei Miao
Associate Advisors: Dr. Caiwen Ding, Dr. Qian Yang
Date/Time: Wednesday, April 24th, 2024, 12:30 pm
Location: HBL 1947 room
WebEx:
Join from the meeting link
https://uconn-cmr.webex.com/uconn-cmr/j.php?MTID=mbab20c5d93769888d1f4adc3a70d298e

Join by meeting number
Meeting number (access code): 2632 104 2737
Meeting password: i3xRV7mN3p4
Abstract
Accurate and efficient modeling of urban mobility, along with the prediction of vehicle and human trajectories, are crucial for ensuring the safety and resilience of intelligent transportation systems and smart cities. However, the dynamic nature of external environments (such as weather conditions, road networks, neighboring vehicles, and passenger behaviors) or intrinsic changes in target data can lead to significant shifts in data distribution. These shifts can invalidate the trained deep learning models and foster overconfidence in model outputs. In this thesis, we introduce both the learning-based and statistical-based methods to address these issues.

First, we design an attention based method for the citywide anomaly event prediction. This method effectively models the spatio-temporal characteristics of urban anomaly events and quantifies the varying impacts of urban mobility on the occurrence of anomaly events. Second, we present an extreme-aware framework to predict the citywide urban mobility under anomalous situations. The proposed framework decomposes the regional urban mobility into spatio-temporally varying regular and extreme dynamics. It minimizes the citywide urban mobility prediction loss under distribution shift. Finally, we introduce a conformal prediction based and Gaussian process regression based framework to quantify the output uncertainty of existing trajectory prediction models (base models). This framework aims to improve the prediction accuracy of the base models and reduce the uncertainty of the predicted trajectories under distribution shift.