[Paper Review] Improving the open cluster census. I. Comparison of clustering algorithms applied to Gaia DR2 data

Emily L. Hunt, S. Reffert|arXiv (Cornell University)|Dec 8, 2020

Stellar, planetary, and galactic studies48 references4 citations

TL;DR

This study compares DBSCAN, HDBSCAN, and Gaussian Mixture Models (GMMs) for detecting open clusters in Gaia DR2 data, finding HDBSCAN most effective due to its superior sensitivity (up to 82%) and ability to recover clusters across varying distances and densities, despite requiring post-processing to reduce false positives. It identifies 41 new open cluster candidates, three within 500 pc, suggesting the Milky Way's open cluster census remains incomplete.

ABSTRACT

The census of open clusters in the Milky Way is in a never-before seen state of flux. Recent works have reported hundreds of new open clusters thanks to the incredible astrometric quality of the Gaia satellite, but other works have also reported that many open clusters discovered in the pre Gaia era may be associations. We aim to conduct a comparison of clustering algorithms used to detect open clusters, attempting to statistically quantify their strengths and weaknesses by deriving the sensitivity, specificity, and precision of each as well as their true positive rate against a larger sample. We selected DBSCAN, HDBSCAN, and Gaussian mixture models for further study, owing to their speed and appropriateness for use with Gaia data. We developed a preprocessing pipeline for Gaia data and developed the algorithms further for the specific application to open clusters. We derived detection rates for all 1385 open clusters in the fields in our study as well as more detailed performance statistics for 100 of these open clusters. DBSCAN was sensitive to 50% to 62% of the true positive open clusters in our sample, with generally very good specificity and precision. HDBSCAN traded precision for a higher sensitivity of up to 82%, especially across different distances and scales of open clusters. Gaussian mixture models were slow and only sensitive to 33% of open clusters in our sample, which tended to be larger objects. Additionally, we report on 41 new open cluster candidates detected by HDBSCAN, three of which are closer than 500 pc. When used with additional post-processing to mitigate its false positives, we have found that HDBSCAN is the most sensitive and effective algorithm for recovering open clusters in Gaia data. Our results suggest that many more new and already reported open clusters have yet to be detected in Gaia data.

Motivation & Objective

• To evaluate and compare the performance of clustering algorithms—DBSCAN, HDBSCAN, and GMMs—for detecting open clusters in Gaia DR2 data.
• To quantify the sensitivity, specificity, precision, and true positive rates of each algorithm across diverse cluster properties.
• To identify limitations in existing cluster detection methods and improve the completeness of the open cluster census in the Milky Way.
• To detect new open cluster candidates, especially nearby and low-surface-brightness objects, using optimized clustering techniques.
• To provide a foundation for future large-scale blind searches using HDBSCAN with post-processing to mitigate false positives.

Proposed method

• Applied DBSCAN, HDBSCAN, and GMMs to Gaia DR2 astrometric data (positions, proper motions, parallaxes) after developing a tailored preprocessing pipeline.
• Used a hybrid approach for DBSCAN with epsilon determined via the ACG method to balance sensitivity and precision.
• Enhanced HDBSCAN by adapting its hierarchical clustering structure to better handle variable cluster densities and scales in Gaia data.
• Optimized GMMs for membership assignment in known clusters, though computational cost limited scalability.
• Evaluated algorithm performance using a sample of 1385 known open clusters, with detailed analysis on 100 representative clusters.
• Applied post-processing to HDBSCAN outputs to filter false positives and validate new cluster candidates.

Experimental results

Research questions

• RQ1 How do DBSCAN, HDBSCAN, and GMMs compare in sensitivity, specificity, and precision for detecting open clusters in Gaia DR2 data?
• RQ2 What are the limitations of each algorithm in recovering clusters of varying distances, sizes, and densities?
• RQ3 Can HDBSCAN detect open clusters missed by DBSCAN and GMMs, particularly in high-extinction or distant regions?
• RQ4 How many new open cluster candidates can be reliably identified using HDBSCAN with post-processing?
• RQ5 To what extent are previously reported open clusters in the MWSC catalogue actually associations or undetectable in Gaia data?

Key findings

• HDBSCAN achieved the highest sensitivity, detecting up to 82% of true positive open clusters, especially across diverse distances and scales.
• DBSCAN detected 50–62% of known clusters with high specificity and precision, but struggled with variable density structures due to its single global epsilon parameter.
• GMMs were the slowest and least sensitive, detecting only 33% of clusters, primarily larger ones, and were unsuitable for large-scale blind searches.
• The study identified 41 new open cluster candidates, three of which are within 500 pc, including one at just 290 pc, indicating significant incompleteness in the local census.
• Several literature clusters from the MWSC catalogue were found to be likely associations or undetectable in Gaia data, suggesting re-evaluation of existing catalogues is needed.
• Post-processed HDBSCAN results show strong promise for future large-scale blind searches, with improved detection of faint and distant clusters.