Robust Cluster Analysis (TCLUST)
Trimmed Robust Cluster Analysis (TCLUST) · Also known as: TCLUST, trimmed clustering, robust clustering, Robust Küme Analizi (TCLUST)
Robust Cluster Analysis is a trimmed model-based clustering method, introduced by García-Escudero and colleagues in 2008, that partitions continuous multivariate data into clusters while resisting the influence of outliers and noise. By setting aside a fraction of the most discordant observations, it keeps the recovered cluster structure from being contaminated by stray points.
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 Robust Cluster Analysis when you want to group continuous, multivariate observations and you suspect the data contain outliers or background noise that would distort ordinary clustering. It assumes continuous multivariate data, requires you to fix the number of clusters k beforehand, and requires you to set a trimming proportion α. A reasonably large sample is needed (at least about 50 observations); below roughly 30 the trimming parameter becomes unreliable because too large a share of observations would be lost. It is less suitable when the outlier ratio exceeds about 0.25, where forcing α high enough can itself damage the cluster structure.
Strengths & limitations
- Resists outliers and contaminating observations that would distort ordinary clustering.
- Impartial trimming lets the method decide which points to exclude rather than relying on a prior outlier rule.
- Model-based formulation captures clusters of differing shape and spread through component covariances.
- The number of clusters k and the trimming proportion α must both be chosen in advance, and results are sensitive to these choices.
- Unreliable in small samples (n < 30), where the share of observations trimmed becomes too high to be trustworthy.
- When the outlier ratio exceeds about 0.25, raising α far enough can itself break the cluster structure.
Frequently asked
What does the trimming proportion α control?
α is the share of observations the method is allowed to discard as outliers. With α = 0.1, the cleanest 90% of the data drive the clustering and the worst-fitting 10% are set aside. Choosing α too low lets outliers contaminate the clusters; choosing it too high can throw away genuine structure.
How is TCLUST different from ordinary k-means or Gaussian mixtures?
Ordinary k-means and Gaussian mixtures assign every point to a cluster, so outliers distort the centres and covariances. TCLUST maximises a trimmed likelihood that leaves a fraction of the worst-fitting points unassigned, so the cluster estimates come only from the clean subset.
Do I have to choose the number of clusters myself?
Yes. Like most partitional methods, TCLUST requires the number of clusters k to be fixed in advance, and a restriction factor on the covariances keeps cluster shapes from degenerating. Monitoring the solution across different k and α values helps justify the final choice.
When should I not use TCLUST?
Avoid it in small samples (below about 30 observations), where the trimming parameter is unreliable and too many points would be lost, and when the outlier ratio is very high (above roughly 0.25), where the trimming needed to clean the data can itself destroy the cluster structure. Robust PCA is a sensible alternative in these cases.
Sources
- García-Escudero, L. A., Gordaliza, A., Matrán, C., & Mayo-Iscar, A. (2008). A General Trimming Approach to Robust Cluster Analysis. The Annals of Statistics, 36(3), 1324-1345. DOI: 10.1214/07-AOS515 ↗
- Riani, M., Cerioli, A., Atkinson, A. C., & Perrotta, D. (2014). Monitoring Robust Regression / Robust Clustering. Statistics and Computing. link ↗
How to cite this page
ScholarGate. (2026, June 1). Trimmed Robust Cluster Analysis (TCLUST). ScholarGate. https://scholargate.app/en/statistics/robust-cluster-analysis
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.
- Cluster-Robust Standard ErrorsStatistics↔ compare
- MM-EstimatorStatistics↔ compare
- Robust Discriminant AnalysisStatistics↔ compare
- Robust PCAStatistics↔ compare
- W-EstimatorStatistics↔ compare