Ensemble HDBSCAN
Also known as: HDBSCAN ensemble clustering, consensus HDBSCAN, multi-run HDBSCAN, cluster ensemble HDBSCAN
Ensemble HDBSCAN runs HDBSCAN multiple times under different hyperparameter settings or data subsamples and combines the resulting partitions into a single stable consensus clustering. Because HDBSCAN is sensitive to its minimum cluster size and minimum samples parameters, pooling multiple runs greatly reduces sensitivity to any single configuration and yields more reproducible cluster assignments on noisy, high-dimensional data.
Key highlights
- Reduces sensitivity to HDBSCAN hyperparameters by averaging over many configurations.
- Inherits HDBSCAN's ability to find arbitrarily shaped, varying-density clusters and identify genuine noise points.
- Produces more reproducible cluster assignments than any single HDBSCAN run.
- Co-association matrix provides a soft membership signal useful for downstream uncertainty analysis.
- Scales well when individual runs are parallelized across cores.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Use Ensemble HDBSCAN when (1) the dataset has complex, varying-density clusters not well captured by spherical-cluster methods like k-means, (2) results need to be robust to HDBSCAN hyperparameter choices, or (3) reproducibility across analysts running slightly different settings is required. It is particularly valuable on high-dimensional data — genomics, text embeddings, sensor streams — where density structure is real but parameter sensitivity is high. Avoid it when computational budget is tight and a single well-tuned HDBSCAN run suffices, or when clusters are roughly spherical and k-means or GMM is faster and simpler. It is not appropriate for labeled data where supervised or semi-supervised methods apply.
Strengths & limitations
- Reduces sensitivity to HDBSCAN hyperparameters by averaging over many configurations.
- Inherits HDBSCAN's ability to find arbitrarily shaped, varying-density clusters and identify genuine noise points.
- Produces more reproducible cluster assignments than any single HDBSCAN run.
- Co-association matrix provides a soft membership signal useful for downstream uncertainty analysis.
- Scales well when individual runs are parallelized across cores.
- Significantly more computationally expensive than a single HDBSCAN run; cost scales with the number of ensemble members.
- The consensus step introduces its own hyperparameter choices (threshold, consensus algorithm) that must be validated.
- Noise handling across runs is non-trivial: points labeled as noise in many runs may be ambiguously placed.
- No widely adopted standard implementation; users typically must assemble the pipeline from HDBSCAN and custom consensus code.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
How many HDBSCAN runs should the ensemble include?
A minimum of 20–50 runs is generally recommended for the co-association matrix to stabilize. Beyond 100 runs, gains in stability are typically marginal while compute cost continues to grow. Monitor co-association matrix convergence as you add runs.
How should noise points be handled in the ensemble?
Treat each run's noise points as singletons when building the co-association matrix — they count as in their own cluster for that run. After consensus, points with consistently low co-association with any group can be designated as stable noise.
Which consensus algorithm works best with HDBSCAN partitions?
Average-link hierarchical clustering on 1 minus the co-association matrix is a common and well-validated choice. Graph-based methods such as spectral clustering on the co-association graph can work better on very large datasets where hierarchical clustering is slow.
How do I choose the consensus threshold?
Treat the threshold as a hyperparameter and select it using an internal validity index (DBCV or silhouette). A threshold around 0.5 is a common starting point, but datasets with very tight or very loose clusters may need different values.
When should I prefer a single HDBSCAN run over the ensemble?
If domain knowledge reliably pins down the hyperparameters, or if computational budget is very limited, a single well-tuned HDBSCAN run is faster and often sufficient. The ensemble pays off most when hyperparameter choice is uncertain or when different analysts must reproduce the same segmentation.
Sources
- 1.McInnes, L., Healy, J., & Astels, S. (2017). hdbscan: Hierarchical density based clustering. Journal of Open Source Software, 2(11), 205.
- 2.Vega-Pons, S., & Ruiz-Shulcloper, J. (2011). A survey of clustering ensemble methods. International Journal of Pattern Recognition and Artificial Intelligence, 25(03), 337–372.
You have read it. What now?
Cite this page
ScholarGate. (2026, June 3). Ensemble HDBSCAN. ScholarGate. https://scholargate.app/machine-learning/ensemble-hdbscan