UAPD: Predicting Urban Anomalies from Spatial-Temporal Data

Urban city environments face the challenge of disturbances, which can create inconveniences for its citizens. These require timely detection and resolution, and more importantly timely preparedness on the part of city officials. We term these disturbances as anomalies, and pose the problem statement: if it is possible to also predict these anomalous events (proactive), and not just detect (reactive). While significant effort has been made in detecting anomalies in existing urban data, the prediction of future urban anomalies is much less well studied and understood. In this work, we formalize the future anomaly prediction problem in urban environments, such that those can be addressed in a more efficient and effective manner. We develop the Urban Anomaly PreDiction (UAPD) framework, which addresses a number of challenges, including the dynamic, spatial varieties of different categories of anomalies. Given the urban anomaly data to date, UAPD first detects the change point of each type of anomalies in the temporal dimension and then uses a tensor decomposition model to decouple the interrelations between the spatial and categorical dimensions. Finally, UAPD applies an autoregression method to predict which categories of anomalies will happen at each region in the future. We conduct extensive experiments in two urban environments, namely New York City and Pittsburgh. Experimental results demonstrate that UAPD outperforms alternative baselines across various settings, including different region and time-frame scales, as well as diverse categories of anomalies. Code related to this chapter is available at: https://bitbucket.org/xianwu9/uapd.