An Experimental Comparison of GPU Techniques for DBSCAN Clustering
Hamza Mustafa, Eleazar Leal, Le Gruenwald · 2019
DBSCAN is a density-based clustering algorithm that is especially useful for finding clusters of arbitrary shapes. As opposed to other clustering techniques, like K-means, it does not require the number of clusters to be specified as an input parameter, and it is highly robust to outliers. However, DBSCAN has a worst-case quadratic time complexity, which makes it difficult to handle large dataset sizes. To address this problem, several works have been proposed that exploit the massive parallelism of GPUs in DBSCAN clustering. Nonetheless, none of these works have been experimentally compared against each other. In this paper, we review the existing GPU algorithms for DBSCAN clustering and conduct the first experimental study comparing these GPU algorithms using three real-world datasets to identify the best performing algorithm. Our results show that CUDA-DClust is the best performing GPU algorithm in terms of execution time and memory requirements.