Significance Measurements for Spatiotemporal Co-occurrences

Abstract

An important aspect of data mining research is the determination of the interestingness of patterns. In classical frequent pattern mining tasks (e.g., shopping basket analysis), the main goal is to identify items (e.g., types of purchased goods) frequently appearing together in an itemset (e.g., shopping cart). Such analyses require an appropriate interestingness measure to assess the strength of relationships among different types of items and to eliminate the spurious itemsets. Measures, such as support, confidence, correlation, and entropy, have been extensively used in many frequent pattern mining algorithms. Spatial and spatiotemporal extensions of frequent pattern mining presents a similar challenge, where the choice of measures may lead to the discovery of inadvisable or uninteresting information depending on the context. Though, unlike traditional frequent pattern mining from binary features, in both spatial and spatiotemporal pattern mining tasks, the spatial or spatiotemporal relationships among items (or instances) are often not explicit. Therefore, it is considered necessary to initially transform the implicit spatial and temporal information to a transaction-like embodiment. In this chapter, we will explore the interestingness measures from the perspective of spatiotemporal co-occurrence relationships appearing among the evolving region trajectories.