Scalable Cluster Analysis of Spatial Events

Abstract

Clustering of massive data is an important analysis tool but also challenging since the data often does not fit in RAM. Many clustering algorithms are thus severely memory-bound. This paper proposes a deterministic density clustering algorithm based on DBSCAN that allows to discover arbitrary shaped clusters of spatio-temporal events that (1) achieves scalability to very large datasets not fitting in RAM and (2) exhibits significant execution time improvements for processing the full dataset compared to plain DBSCAN. The proposed algorithm's integration with interactive visualization methods allows for visual inspection of clustering results in the context of the analysis task; several alternatives are discussed by means of an application example about traffic data analysis.