ATD: Deep Learning on Anomaly Detection for Human Dynamics and Hazard Response

University Of Massachusetts Amherst, Amherst MA

Investigators

Abstract

The project aims to investigate mathematical models that can provide a deeper understanding of human risk response. The analysis of human movement patterns in space and time, at various levels of granularity, holds crucial importance in fields such as transport management, healthcare, and threat detection. Over the past decade, the proliferation of smartphones and Global Positioning System (GPS)-enabled devices has granted us unprecedented access to vast amounts of location data, timestamped with high precision. Leveraging this data, the project will focus on anomaly detection to identify unexpected or significantly different behaviors within observed mobility datasets. The outcomes of this research will prove valuable in detecting both natural and human-induced hazard situations, enabling more effective responses. The research will have numerous practical applications, including pandemic contact tracing and spread modeling, hazard evacuation planning, and digital footprint tracking. Moreover, the project will provide training opportunities for students from underrepresented groups in STEM fields. The project aims to develop a new deep learning framework implemented on a Geographic Information System (GIS) platform. This framework will advance big spatiotemporal data analytics, specifically in anomaly detection, and quantify human mobility dynamics concerning hazard response behaviors. The research will focus on three main objectives: 1. Develop an agent-based machine learning framework, Markov Decision Process - Inverse Reinforcement Learning - Generative Adversarial Network (MDP-IRL-GAN), to detect anomalies in individual movement dynamics. 2. Model hazard response by analyzing individual movement dynamics using the proposed machine learning framework, while identifying the key factors influencing decision-making in response to hazards. 3. Detect changes and anomalies in spatiotemporal patterns of group dynamics by employing a newly-designed multi-resolution graph neural network (MA-GNN). The results will contribute to the efficiency of anomaly detection, the accuracy of traffic forecasting, and a deeper comprehension of human risk response behaviors. 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.