Anomaly Detection for Shared Mobility

Shared mobility systems, such as bike-sharing networks, have gained significant traction in urban transportation due to their environmental and economic benefits. Understanding anomalies within these systems is critical for improving operational efficiency, user satisfaction, and system resilience. Current state-of-the-art approaches to anomaly detection in shared mobility often focus on user behavior or system-level metrics. However, these methods frequently lack interpretability and fail to address the need for fast, unsupervised techniques, as labeled data is typically unavailable in this domain. This thesis addresses these limitations by proposing a comprehensive approach to anomaly detection in shared mobility systems. The approach integrates diverse data sources, including bike-sharing data, weather data, and mass transit data, to provide a richer contextual understanding of anomalies. Machine learning techniques, such as Isolation Forest, are utilized alongside feature selection methods like Depth-based Isolation Forest Feature Importance (DIFFI) to identify atypical patterns in usage that may indicate operational inefficiencies or unusual user behavior. Key contributions of this work include the development of spatiotemporal, environmental, and transit-related features, the application of interpretable anomaly detection models, and an in-depth analysis of the relationships between anomalies and external factors. These findings offer actionable insights for shared mobility operators, enabling enhancements in system performance and user experience.