ICAD: A Self-Supervised Autoregressive Approach for Multi-Context Anomaly Detection in Human Mobility Data.

Area of Science:

• Computational Social Science
• Data Science
• Artificial Intelligence

Background:

• Abnormal human mobility patterns can indicate emergencies or health risks, necessitating effective detection methods.
• Current anomaly detection approaches often miss fine-grained temporal details and lack interpretability regarding contributing factors.

Purpose of the Study:

• To introduce ICAD (Interpretable Component-wise Anomaly Detection), a novel self-supervised model for detecting spatial and temporal anomalies in human mobility.
• To enhance the interpretability of anomaly detection by providing component-wise anomaly scores.

Main Methods:

• Developed ICAD, a self-supervised autoregressive model trained on normal visit sequences using a next-visit prediction objective.
• Implemented component-wise scoring, including a top-k deviation metric for spatial anomalies and a relative mode-based scoring for temporal anomalies.

Main Results:

• ICAD effectively detects both spatial and temporal anomalies at the visit-level.
• The model demonstrates superior performance compared to existing methods in both visit-level and agent-level anomaly detection on synthetic data.

Conclusions:

• ICAD offers a more interpretable and comprehensive approach to human mobility anomaly detection.
• The model's ability to identify fine-grained spatiotemporal deviations has significant implications for public safety and healthcare applications.