Ensemble K-means
Ensemble K-means Clustering (Consensus Clustering) · Also known as: consensus K-means, K-means ensemble clustering, cluster ensemble with K-means, EKM
Ensemble K-means runs K-means clustering many times under varied initializations, random seeds, or feature subsets, then aggregates the resulting partitions into a single consensus assignment. This approach reduces K-means' well-known sensitivity to initialization and produces more stable, reproducible clusters than any single run.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use Ensemble K-means when a single K-means run gives unstable or irreproducible results across repeated analyses, when you are uncertain about the number of clusters k, or when the dataset has complex or elongated clusters that a single K-means initialization often misses. It is especially valuable in genomics, image segmentation, and customer segmentation where reproducibility matters. Do not use it as a substitute for domain-driven cluster validation: always interpret clusters substantively. Avoid it when computational budget is very tight (B runs cost B times the single-run time) or when the data has very high dimensionality without dimensionality reduction, as the co-association matrix becomes noisy.
Strengths & limitations
- Dramatically more stable and reproducible clusters than a single K-means run.
- Agnostic to label permutation across runs, avoiding the correspondence problem.
- Can aggregate runs with different k values, giving implicit robustness to the number-of-clusters choice.
- The co-association matrix provides a soft similarity measure useful for downstream analysis.
- Parallelizable: each of the B K-means runs is independent and can run concurrently.
- Computational cost scales linearly with the number of ensemble members B.
- Storing and processing the n-by-n co-association matrix is memory-intensive for large n.
- Final consensus clustering still requires choosing a method and its parameters (e.g., number of clusters).
- Does not resolve the fundamental assumption of K-means (spherical, similarly sized clusters) unless combined with diverse base clusterers.
Frequently asked
How many ensemble members B should I use?
A common default is 50–100 runs. Increase B until the co-association matrix stabilizes — that is, the consensus cluster assignments no longer change when you add more runs. Fewer than 20 runs rarely provide meaningful stability gains.
How do I choose the number of clusters k?
Run the ensemble over a range of k values and inspect the consensus cumulative distribution function (CDF) plots. The k at which the CDF shows the most abrupt change (highest delta) indicates the most distinct clustering. The consensus approach is specifically designed to make this selection more reliable than the elbow method on a single K-means run.
Does Ensemble K-means solve K-means' assumption of spherical clusters?
Not fully. If all ensemble members use standard K-means with Euclidean distance, the aggregate still inherits a preference for convex, roughly spherical clusters. For non-convex or arbitrarily shaped clusters, combine ensemble clustering with kernel K-means or replace some base clusterers with DBSCAN.
What is the difference between Ensemble K-means and just running K-means many times and picking the best?
Picking the best single run by lowest inertia is a greedy deterministic choice; it discards the information in all other runs. Ensemble K-means aggregates all runs into a co-association matrix, capturing the probabilistic co-occurrence structure across runs, which is strictly more informative than any single winning partition.
Is this suitable for very large datasets?
The co-association matrix is n-by-n, so memory becomes a bottleneck beyond roughly 50,000 points. For larger datasets, use sparse co-association approximations, mini-batch K-means as base clusterer, or subsampling strategies such as the consensus clustering variant of Monti et al.
Sources
- Strehl, A. & Ghosh, J. (2002). Cluster ensembles — a knowledge reuse framework for combining multiple partitions. Journal of Machine Learning Research, 3, 583–617. link ↗
- Monti, S., Tamayo, P., Mesirov, J. & Golub, T. (2003). Consensus clustering: a resampling-based method for class discovery and visualization of gene expression microarray data. Machine Learning, 52, 91–118. DOI: 10.1023/A:1023949509487 ↗
How to cite this page
ScholarGate. (2026, June 3). Ensemble K-means Clustering (Consensus Clustering). ScholarGate. https://scholargate.app/en/machine-learning/ensemble-k-means
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Ensemble Gaussian Mixture ModelMachine learning↔ compare
- K-meansMachine learning↔ compare
- Semi-supervised K-meansMachine learning↔ compare