Explainable K-Means
Explainable K-Means Clustering · Also known as: ExKMC, interpretable k-means, decision-tree k-means, explainable clustering
Explainable K-Means is a post-hoc and in-model interpretability approach to standard K-Means clustering that replaces or approximates cluster assignments with a small axis-aligned decision tree. Each leaf of the tree corresponds to one cluster, and every data point is assigned to a cluster by following a simple sequence of threshold rules on individual features — making cluster membership fully transparent and human-readable.
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 Explainable K-Means when the goal is to segment data into interpretable groups whose membership rules must be communicated to non-technical stakeholders — for example, customer segments described by simple business rules, or patient subgroups defined by clinical thresholds. It is well-suited to tabular data with continuous features and moderate k (2–10). Avoid it when the number of clusters is large (k > 20), when the true cluster shapes are highly non-spherical (where DBSCAN or GMM is preferable), or when maximising clustering quality is more important than transparency.
Strengths & limitations
- Cluster membership is described by simple if-then threshold rules that are immediately understandable by domain experts and non-technical audiences.
- Approximation guarantees exist: theoretical bounds on how much clustering quality is lost relative to standard K-Means are established in the literature.
- Applies to any K-Means solution as a post-hoc step, so it integrates with existing clustering pipelines without replacing them.
- Decision-tree structure enables fast assignment at inference time — no centroid distance computation is needed.
- Supports audit and regulatory requirements that demand explainable automated decisions.
- Axis-aligned splits cannot perfectly capture diagonal or curved K-Means boundaries, so some clustering quality is always sacrificed.
- The approximation cost gap grows with k and dataset complexity — large k values produce trees that are either deep or inaccurate.
- The greedy tree-construction heuristic does not guarantee the globally optimal explainable partition.
- Performance degrades when true clusters are elongated, overlapping, or non-convex.
Frequently asked
How does Explainable K-Means differ from just running a decision tree on K-Means labels?
A post-hoc decision tree fitted on K-Means labels optimises classification accuracy (correct label prediction) and may not minimise clustering cost. Explainable K-Means directly minimises the within-cluster sum of squares under the tree's induced partition, producing approximation guarantees that a generic classifier does not provide.
How many clusters (k) is too many for the method to remain interpretable?
In practice, interpretability begins to degrade beyond k = 8–10 because the tree grows deep enough to lose its intuitive readability. Theoretical guarantees still hold, but communicating the result becomes difficult. For large k, consider hierarchical approaches or topic models instead.
Does the method require rerunning K-Means, or can it work on any existing clustering?
Explainable K-Means uses the K-Means centroids as a reference target. It can work post-hoc on any existing K-Means solution — you do not need to rerun K-Means. However, the centroids are required to compute the approximation cost.
What is a typical approximation cost ratio, and how large a quality loss is acceptable?
Empirical studies report approximation ratios of 1.1–2.0 on real datasets, meaning the tree-based clustering costs 10–100% more than the K-Means optimum. Whether this is acceptable depends on the application; in domains where transparency is mandatory, even a 2x cost increase is often justified.
Can Explainable K-Means handle categorical features?
The method is designed for continuous features because K-Means uses Euclidean distances. Categorical features must be numerically encoded first (e.g., ordinal encoding or binary indicators), and the resulting threshold rules on encoded values must be interpreted carefully.
Sources
- Dasgupta, S., Frost, N., Moshkovitz, M., & Rashtchian, C. (2020). Explainability of k-Means Clustering. Proceedings of the 37th International Conference on Machine Learning (ICML), PMLR 119. link ↗
- Moshkovitz, M., Dasgupta, S., Rashtchian, C., & Frost, N. (2020). Explainable k-Means and k-Medians Clustering. Proceedings of the 37th International Conference on Machine Learning (ICML), PMLR 119. link ↗
How to cite this page
ScholarGate. (2026, June 3). Explainable K-Means Clustering. ScholarGate. https://scholargate.app/en/machine-learning/explainable-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.
- DBSCANMachine learning↔ compare
- Decision TreeMachine learning↔ compare
- Hierarchical ClusteringMachine learning↔ compare
- K-Means ClusteringMachine learning↔ compare
- Random ForestMachine learning↔ compare