CRII: III: Learning Spatiotemporal Impacts of Text-enriched Traffic Events with Injection of Interpretability from Graph Neural Networks and Physics-Informed Machine Learning

Predicting the future spatiotemporal impacts and underlying cascading patterns of the traffic incidents that occur within the transportation systems could offer considerable benefits to society. As traffic incidents are responsible for approximately 25% of all traffic delays, costing the United States economy an estimated $87 billion annually, an accurate and explainable predictive solution can help mitigate these devastating spatiotemporal impacts by enabling quicker response times from emergency services, more effective traffic management, and better public advisories. Advanced analytics and machine learning algorithms can analyze real-world data to forecast the severity and duration of spatiotemporal events that occurred within the Cyber-Physical-Social Systems, thereby aiding in the allocation of resources and reducing both human and economic losses. The primary innovation of this project will be its ability to learn the complex relationship between traffic incidents and how their impact will cascade among the geometric structure of transportation systems and extract understandable rules and patterns for decision-makers.

The objective of this project is to gather and analyze extensive traffic data from various heterogeneous platforms to learn more representative embeddings of traffic incidents. The goal is to create a range of interpretable graph mining and machine learning techniques that can enhance understanding of the cascading impacts not only in transportation systems but also across heterogenous cyber-physical-social systems. This project will develop 1) a novel multimodal representation learning approach combining machine learning and graph mining to learn the text-enriched embedding of the traffic events from heterogenous social/news media and transportation networks, 2) novel spatiotemporal data mining and transformer-based graph neural network methods to detect transportation events and their cascading impacts on heterogeneous transportation networks, 3) a new physics-informed machine learning solution to model the cascading impacts with the Korteweg-de Vries (KdV) equation and inject model interpretability while forecasting the traffic events in specific scenarios. The proposed physics-informed machine learning solutions could decouple the model's dependencies on large datasets so that the transportation systems in less represented/rural areas could swiftly infer accurate future statuses when failure happens. These advancements will make significant contributions to the fields of explainable graph mining, physics-informed machine learning, and spatiotemporal event analysis. The expected outcomes will enrich the understanding of the nature of traffic events in complex transportation networks.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

Award Number: 2551684
Principal Investigator: Kaiqun Fu
Funds Obligated: $168,912
State: TX