Density and Frequency-Aware Cluster Identification for Spatio-Temporal Sequence Data

Various wireless sensors and devices keep collecting data for their environments or owners. Such collected data are given in the form of spatio-temporal sequence data which are a sequence of data elements with spatial information and timestamp. Data clustering is useful in finding inherent underlying structures, natural or interesting groups in a collection of data. This paper proposes a new clustering method for spatio-temporal sequence data with respect to density and frequency. Density is a notion about how densely data elements are in a local region, and frequency is a notion about how many times sequences pass through a local region. The proposed method identifies three types of clusters: high density and high frequency clusters, high density and low frequency clusters, and low density and high frequency clusters. It first augments the data set by inserting dummy data elements for capturing frequency distribution in sparse local regions. Then it computes the densities and frequency for data elements and the frequencies for dummy data elements. It partitions data elements into the high density-high frequency data set, high density-low frequency data set, and low-density-high frequency data set. It clusters each data set individually using the clustering procedures that are similar to DBSCAN, which is a density-based clustering algorithm. The proposed method had been applied to the six spatial-temporal GPS sequence data sets for wildlife movements. The experiment results were compared with the results from DBSCAN and analyzed in terms of the number and characteristics of discovered clusters.