Theoretically-Efficient and Practical Parallel DBSCAN

TL;DR

This paper introduces new parallel algorithms for Euclidean DBSCAN that match the work bounds of sequential algorithms and are highly parallel, significantly improving practical performance and scalability.

Contribution

It presents the first parallel algorithms for exact and approximate DBSCAN that achieve optimal work bounds and low depth, bridging the gap between theory and practice.

Findings

Outperform existing parallel DBSCAN by up to several orders of magnitude.

Achieve speedups of up to 33x over the best sequential algorithms.

Demonstrate practical efficiency on a 36-core machine with hyper-threading.

Abstract

The DBSCAN method for spatial clustering has received significant attention due to its applicability in a variety of data analysis tasks. There are fast sequential algorithms for DBSCAN in Euclidean space that take $O(n\log n)$ work for two dimensions, sub-quadratic work for three or more dimensions, and can be computed approximately in linear work for any constant number of dimensions. However, existing parallel DBSCAN algorithms require quadratic work in the worst case, making them inefficient for large datasets. This paper bridges the gap between theory and practice of parallel DBSCAN by presenting new parallel algorithms for Euclidean exact DBSCAN and approximate DBSCAN that match the work bounds of their sequential counterparts, and are highly parallel (polylogarithmic depth). We present implementations of our algorithms along with optimizations that improve their practical performance. We perform a comprehensive experimental evaluation of our algorithms on a variety of datasets and parameter settings. Our experiments on a 36-core machine with hyper-threading show that we outperform existing parallel DBSCAN implementations by up to several orders of magnitude, and achieve speedups by up to 33x over the best sequential algorithms.