Situation Estimation and Prediction in Spatio-Temporal Data Streams

Situation recognition has been a major challenge in most domains for several decades. With the recent emergence of rapid data dissemination platforms like social media, blogs, and a push towards an Internet of Things, the amount of data about multiple facets of daily life has exploded. This presents an unprecedented opportunity to harness these data streams to determine situations in space and time.;There are several challenges inherent in this goal. The data streams may originate from traditional as well as non-traditional sources. As such, these may manifest remarkable di- versity in the media type and the granularity at which data are observed. Non-traditional sources like Twitter, Pinterest and micro-blogs allow virtually no control on when and where data should be sensed. One has no control over where to deploy these sensors in order to maximize coverage in space and time. The uncertainty associated with these data streams might not be known in advance. There is also the issue of how reliable the data might be, especially the one crowd-sourced from non-traditional sources.;This work aims to develop a data-driven platform that allows application developers to use heterogeneous spatio-temporal data streams to estimate the underlying situation of interest and perform short term prediction on those. We introduce data structures to handle uncertainty in data which also facilitates a probabilistic treatment of estimation and prediction methods. Probabilistic approach also lets us handle missing values and data coverage issues by marginalizing the unknown spatio-temporal elements.;The proposed framework uses context defined by the user to specify different models for different context. This is helpful in modeling estimation and prediction procedures as this does not adhere to a one-model-fits-all approach. There are also constructs to learn the relationships between observations and situations, and to characterize the noise associated with the observation data stream.;We propose how one may estimate and predict recurrent situations along with incorporating the impacts of external events and factors which might affect the situation. As an application of this framework, we discuss how one may estimate the traffic speeds on various freeways, in the presence of disrupting factors like accidents and public events. We also apply the framework to estimate the popularity of the Democrats as compared to that of Republicans for the 2012 US Presidential elections. A third application predicts crimes in the City of Chicago based on previously recorded crimes.