A general framework for mining spatial and spatio-temporal object association patterns in scientific data

Advances in computational sciences and data collection techniques have resulted in the accumulation of huge amounts of spatial or spatio-temporal data in a wide range of scientific disciplines, such as bioinformatics, astrophysics, meteorology, and computational fluid dynamics. As a result, data mining---the process of discovering hidden and useful information in datasets---has been employed to facilitate the understanding of important phenomena in such disciplines. Many approaches have been proposed to analyze spatial or spatio-temporal data. However, they often suffer from several major limitations. First, they often model spatial entities as points. However, this leads to a loss in information since the geometric properties of such entities (or features) can play an important role in many scientific applications. Second, they lack effective schemes to model the diversity of spatial or spatio-temporal relationships among features. Modeling such relationships are key to understanding the evolutionary behavior of features in many scientific domains. Finally, they are often not cognizant of domain knowledge when modeling these relationships. This can limit the usefulness of the data mining process and inhibit our ability to effectively reason about important scientific phenomena.; In this dissertation, we present a general and modularized framework to address these limitations when mining spatial or spatio-temporal scientific data. We propose different representation schemes to model the geometric properties of spatial entities. We define Spatial Object Association Patterns (SOAPs) to characterize a variety of relationships among entities. Furthermore, we introduce SOAP episodes to capture the evolutionary nature of such relationships. In addition, we propose multiple reasoning strategies to infer important events based on SOAPs or SOAP episodes. We empirically demonstrate the efficacy of this framework on applications originating from the following scientific disciplines: bioinformatics, computational molecular dynamics, and computational fluid dynamics. Our results show that the framework can discover meaningful and important spatial or spatio-temporal patterns. We also demonstrate that the proposed reasoning strategies can make meaningful inferences on important phenomena in such scientific disciplines. Finally, through such applications, we have empirically shown the potential of employing the proposed framework to realize automated or semi-automated data analysis in different scientific disciplines.