Pattern discovery from spatiotemporal data

Abstract

In many applications that track and analyze spatiotemporal data (i.e., moving objects), the movement of objects exhibits regularities. The subject of this thesis is the discovery of three types of such regularities from spatiotemporal sequences; periodic patterns; spatiotemporal sequential patterns, and spatiotemporal collocation episodes. The first problem is motivated by the fact that many movements obey periodic patterns; the objects follow the same routes (approximately) over regular time intervals. For example, people wake up at the same time and follow more or less the same routes to their work everyday. The approximate nature of such patterns renders existing models, based on symbol sequences, inadequate. We define the problem of mining periodic patterns in spatiotemporal data and propose appropriate algorithms for its solution. In addition, we adapt our mining techniques for two interesting variants of the problem: (i) the extraction of periodic patterns that are frequent in a continuous sub-interval of the whole history, and (ii) the discovery of periodic patterns in which some instances may be shifted or distorted. We present a comprehensive experimental evaluation showing the effectiveness and efficiency of the proposed techniques. The second problem is the discovery of spatiotemporal sequential patterns which are routes frequently followed by the objects in irregular time. The challenges to address are the fuzziness of the locations in a long input sequence and the identification of non-explicit pattern instances. We define pattern elements as spatial regions around frequent line segments. Our method first transforms the original sequence into a list of sequence segments, and detects frequent regions heuristically. Then, we propose algorithms that find patterns by employing a new substring tree structure and an improvement upon the Apriori technique. An experimental study demonstrates the effectiveness and efficiency of our approach. The third type of patterns captures the inter-movement regularities among different types of objects. Given a collection of trajectories of moving objects with different classes, (e.g., pumas, deers, vultures, etc.), we extract collocation episodes in them, (e.g., if a puma is moving near a deer, then a vulture will also move close to the same deer with high probability within the next 3 minutes). Although a lot of work in the spatial collocation retrieval and the discovery of frequent episodes in temporal data has been done, to our best knowledge this is the first work that combines the two concepts to detect interesting information from spatiotemporal data sequences. We formally define the problem of mining collocation episodes and provide a two-phase framework to solve it. The first phase applies a hash-based technique to identify pairs of objects that move closely in some time intervals. In the second phase, we provide two algorithms and an optimization technique to discover the collocations. We empirically evaluate the performance of the methods using synthetically generated data that emulate the real-world object movements.