Mobility-Based Anomaly Detection

Open Access
- Author:
- Xin, Yanan
- Graduate Program:
- Geography
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- June 15, 2020
- Committee Members:
- Alan M Mac Eachren, Dissertation Advisor/Co-Advisor
Alan M Mac Eachren, Committee Chair/Co-Chair
Manzhu Yu, Committee Member
Justine Blanford, Committee Member
Xiaolong (Luke) Zhang, Outside Member
Cynthia Ann Brewer, Program Head/Chair - Keywords:
- Mobility Analysis
Anomaly Detection
Spatial Data Mining
Deep Learning
Machine Learning - Abstract:
- Mobility data are proliferating at an unprecedented rate due to the ubiquitous GPS sensing and tracking. The increased availability of mobility data gives rise to numerous applications ranging from urban traffic monitoring to participatory environmental sensing. Detecting anomalies observed in mobility data (specified here as mobility-based anomaly detection) has attracted significant attention from researchers and practitioners in various fields due to its significant real-world impact. For example, traffic anomalies are used for traffic accident monitoring, and anomalies in environmental mobile sensing data are used to signal potential natural hazards. Despite a large number of studies available on mobility-based anomaly detection, many of the studies are conducted in distinctly different fields and have not been examined under a unified framework. This dissertation provides a systematic investigation of mobility-based anomaly detection to fill this knowledge gap. I propose a taxonomy of mobility-based anomaly detection to organize the existing relevant studies into three categories based on the source and target attributes of mobility data used in the anomaly detection process: (1) utilizing mobility attributes as both source and target in anomaly detection (mobility to mobility anomaly detection), (2) utilizing mobility attributes as the source and non-mobility attributes as the target (mobility to non-mobility anomaly detection), and (3) utilizing non-mobility attributes as the source and mobility attributes as the target (non-mobility to mobility anomaly detection). Following the taxonomy, three individual studies are presented, with each providing an example for one of the three categories. The first study (an example of mobility to mobility anomaly detection) identifies anomalous patterns of shared dockless e-scooters using an unsupervised deep learning approach. The second study (an example of mobility to non-mobility anomaly detection) detects anomalies in crowdsourced radiation measurements. The third study (an example of non-mobility to mobility anomaly detection) models the atypical event travel patterns of football fans using geolocated tweets. The three studies develop new methods in addressing the challenges of mobility-based anomaly detection and provide insights into the specific application domain. The dissertation provides one of the first systematic efforts to address mobility-based anomaly detection generally and highlights challenges and opportunities for future research.