HPDBSCAN: Highly Parallel DBSCAN
Published in Proceedings of the Workshop on Machine Learning in High-Performance Computing Environments (MLHPC 15), pp. 1–10, 2015
DBSCAN is a widely used density-based clustering algorithm, but its inherently sequential design limits scalability to large datasets. We introduce HPDBSCAN, a highly parallel, distributed-memory reformulation of DBSCAN that partitions the data across compute nodes while preserving the algorithm’s clustering semantics. HPDBSCAN achieves substantial speedups over existing implementations, enabling density-based clustering of large-scale scientific datasets on modern HPC systems.
Recommended citation: Götz, M., Bodenstein, C., & Riedel, M. (2015). "HPDBSCAN: Highly Parallel DBSCAN." In: Proceedings of the Workshop on Machine Learning in High-Performance Computing Environments (MLHPC 15), pp. 1–10.
Download Paper