Robust Hierarchical Clustering
Also known as: robust agglomerative clustering, outlier-resistant hierarchical clustering, robust linkage clustering, RHC
Robust hierarchical clustering extends classical agglomerative or divisive hierarchical clustering by replacing sensitive distance measures and linkage criteria with outlier-resistant alternatives, preserving cluster structure even when data contain anomalous observations or heavy-tailed distributions.
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 hierarchical clustering when your multivariate data may contain outliers, measurement errors, or heavy-tailed variation that could distort classical hierarchical clustering results. It is particularly suitable for exploratory analysis where the number of clusters is unknown and you want both a partition and a visual representation of cluster hierarchy. It works well with continuous variables, though robust distance measures for mixed data exist. Do NOT use it when data are clean and Gaussian — classical hierarchical clustering is sufficient and faster. Avoid it when n is very large (above a few thousand) because agglomerative algorithms are O(n^2) in memory and time; partitioning-based methods such as robust k-means or CLARA scale better. It is also not appropriate as a confirmatory test of a pre-specified cluster structure — use a model-based approach such as robust mixture modeling for that.
Strengths & limitations
- Produces a full dendrogram showing the complete merge hierarchy, aiding interpretation of the number and nesting of clusters.
- Robust distance measures based on MCD or similar estimators substantially reduce the influence of outliers and leverage points.
- No need to specify the number of clusters in advance; the cut can be chosen post hoc from the dendrogram.
- Outliers are naturally revealed as late-joining singletons at large heights, providing an automatic contamination diagnostic.
- Compatible with domain-driven choice of linkage (average, complete, Ward-like) for different cluster shapes.
- Agglomerative algorithms have O(n^2) time and memory complexity, making them impractical for large datasets.
- MCD-based robust distance estimation itself requires sufficient sample size relative to the number of variables (n >> p).
- Merges are irreversible: an early erroneous merge cannot be undone, which can propagate errors through the dendrogram.
- The choice of linkage criterion still influences results and there is no universally optimal choice.
- Robust methods reduce but do not eliminate sensitivity to extreme violations of assumptions.
Frequently asked
How does robust hierarchical clustering differ from classical hierarchical clustering?
Classical hierarchical clustering computes distances using the sample mean and covariance matrix, which are highly sensitive to outliers. Robust hierarchical clustering replaces these with outlier-resistant estimates — most commonly the Minimum Covariance Determinant estimator — and may also use robust linkage criteria, so that a small number of anomalous observations cannot distort the entire cluster solution.
Which linkage method should I use with robust distances?
Average linkage (UPGMA) and complete linkage are generally preferred over single linkage because single linkage is prone to chaining, where clusters elongate by adding one observation at a time. Ward's criterion can also be adapted to robust distances. When outliers are a major concern, trimmed-linkage variants that ignore the most extreme distances between clusters offer additional protection.
Can robust hierarchical clustering handle mixed data types?
It can if appropriate robust distance measures for mixed data are used — for example, robust extensions of Gower's distance. However, most readily available implementations focus on continuous variables. For mixed data, robust model-based clustering or robust latent class analysis may be more principled alternatives.
What sample size do I need?
MCD-based robust distance estimation requires substantially more observations than variables (n >> p). A common practical guideline is n >= 5p at minimum. If p is large relative to n, consider dimensionality reduction first or use regularized robust covariance estimators.
How do I determine the number of clusters?
Inspect the dendrogram for large jumps in merge height, which suggest natural cluster boundaries. Complement this with quantitative criteria computed on robust distances: the silhouette coefficient, the gap statistic, or the Calinski-Harabasz index. Consistency across multiple criteria strengthens the chosen solution.
Sources
- Kaufman, L. & Rousseeuw, P. J. (1990). Finding Groups in Data: An Introduction to Cluster Analysis. Wiley. ISBN: 978-0471878766
- Garcia-Escudero, L. A., Gordaliza, A., Matran, C. & Mayo-Iscar, A. (2010). A review of robust clustering methods. Advances in Data Analysis and Classification, 4(2–3), 89–109. DOI: 10.1007/s11634-010-0064-5 ↗
How to cite this page
ScholarGate. (2026, June 3). Robust Hierarchical Clustering. ScholarGate. https://scholargate.app/en/statistics/robust-hierarchical-clustering
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 AnalysisStatistics↔ compare
- Hierarchical ClusteringMachine learning↔ compare
- Mixture ModelingStatistics↔ compare
- Multidimensional ScalingStatistics↔ compare
- Robust K-means ClusteringStatistics↔ compare