Detecting and Tracking Spatio-temporal Clusters with Adaptive History Filtering

Abstract

This paper addresses the problem of detecting and tracking moving clusters in spatio-temporal data sets. Spatio-temporal data sets contain data elements that move in space over time. Traditional data clustering algorithms work well on static data sets that contain well separated clusters. When traditional techniques are applied to spatio-temporal data they breakdown when the moving data elements intersect the space occupied by elements from another cluster. The goal of this work is to improve the accuracy of traditional data clustering algorithms on spatio-temporal data sets. Many clustering algorithms create clusters based on the distance between the elements. We extend this distance measure to be a function of the position history of the elements. We show through a series of experiments that the use of the history based distance measures greatly improves the performance of existing data clustering algorithms on spatio-temporal data sets. In random data sets we achieve up to a 90% improvement in cluster accuracy. To evaluate the clustering algorithms we created 102 spatio-temporal data sets. We also defined a set of metrics that are used to evaluate the performance of the clustering algorithms on the spatio-temporal data sets.