Anomaly Detection Across Multi-scale Temporal Data Streams for Human Behavior Modeling

This research focuses on developing computational models for human behavior from multi-scale temporal data to detect anomalous behavior, evaluating car driving behavior as a case study. Human behavioral patterns capture frequent or repeated behaviors of users in the data. Here, the user data can be multi-scale from devices, computer networks, or even vehicle driving data. Human behavioral patterns can be associated with identifying anomalies which are precursors or even indicators of impending or ongoing unexpected behavior. We aim to address the discovery of anomalous human behavioral patterns in the driving domain. We present time series-based anomaly detection utilizing car telematics data, eye tracking gaze distraction data, and health vital statistics data to provide a comprehensive view of the driver behavioral patterns. We analyze different scales and resolutions of time and the anomalous variations and their intensities in the data streams. Our results indicate that each of the heterogeneous temporal data streams of telematics data, eye tracking data, and driver vital health data can individually indicate anomalies in driving states varying across safe to unsafe driving state. However, gaze data is more representative of the anomalies than health and telematics data in individual streams. In general, we also found that combinations of all data streams are useful. We are also able to supplement the anomalous state information through the combination and overlap of anomalies in the three data streams.