Large-scale Cluster Behavior Anomaly Detection

With the development of information technology and network technology, more and more people’s behaviors and trajectories are recorded conveniently by tracing equipment like GPS and mobile phones, which makes observe and do research on the society possible. Moreover, with the data collected by these devices, we are able to study the new field of human dynamics.The sudden crises occur or evolve more disorderly in large cities because of the large population. The difficulty of monitoring and dealing with these unexpected crises rises correspondingly. Consequently, more intelligent systems are needed to aid the crisis management and the emergency response. We can monitor the status of the society and adjust timely through the analysis of the temporal and spatial distribution of the population. Once an abnormal behavior of a cluster is accurately predicted, we can prepare in advance and take appropriate measures beforehand.In this thesis, we investigate the large-scale collective behaviors through the analysis of the spatio-temporal data. Considering that the outlier detection on collective behavior is an interesting and practical problem, we put our focus on the detection of abnormal events. A framework for outlier detection on collective behaviors is proposed. Meanwhile, we also proposed several methods to extract features from data gathered in a period of time. The experimental results by using this framework on artificial simulation data further validate the reliability and effectiveness of our methods. Similar results are also obtained by employing real data collected by sensors spread on the roads in Twin Cities.