Distributed Density Peak Clustering of Trajectory Data on Spark

Abstract

With the widespread use of mobile devices and GPS, trajectory data mining has become a very popular research field. However, for many applications, a huge amount of trajectory data is collected, which raises the problem of how to efficiently mine this data. To process large batches of trajectory data, this paper proposes a distributed trajectory clustering algorithm based on density peak clustering, named DTR-DPC. The proposed method partitions the trajectory data into dense and sparse areas during the trajectory partitioning and division stage, and then applies different trajectory division methods for different areas. Then, the algorithm replaces each dense area by a single abstract trajectory to fit the distribution of trajectory points in dense areas, which can reduce the amount of distance calculation. Finally, a novel density peak clustering-based method (E-DPC) for Spark is applied, which requires limited human intervention. Experimental results on several large trajectory datasets show that thanks to the proposed approach, runtime of trajectory clustering can be greatly decreased while obtaining a high accuracy.