Detecting and classifying anomalous behavior in spatiotemporal network data

We investigate different models for detecting and classifying important geopolitical events in high-frequency spatiotemporal network data. Building on previous empirical work on the network response to real-world events, our goal is to develop a generative model that can identify the time, location, and nature of different emergency and non-emergency events. As a testbed for these models, we use a large dataset containing billions of anonymized mobile phone calls and text messages from Afghanistan, and associated metadata on several known important geopolitical events. We find that simple and scalable time-series models of geographically aggregated call volume can accurately identify the onset of major events when the approximate time and location of the event is known. However, such models ignore the network structure in the data, and are not well suited to spatial localization. Preliminary results from dynamic matrix factorization models, which generatively model network structure, indicate a promising area for future work.