Explainable Clustering for Interpretable Unsupervised Learning

This direction is split into other directions: dimension-dependent explainability cost. Its last version stays readable and citable here.

Version 1 of 165 papers66 conceptsfrozen 2026-10-04

Traditional clustering is a "black box" operation. When an algorithm like $k$-means partitions a dataset, it typically provides a list of cluster assignments—"Point A belongs to Cluster 1"—without explaining the underlying logic. In high-stakes domains like medicine, finance, or law, these opaque groupings are difficult to trust or act upon.

The research direction of explainable clustering seeks to replace or augment these numerical assignments with human-readable rules, ensuring that every cluster has a concise, feature-based justification.

The Problem with Distance-Based Grouping

Most clustering algorithms rely on complex, multi-dimensional distance metrics. As noted by Konstantin Makarychev and Liren Shan, understanding why a point belongs to a specific cluster in $k$-means requires calculating distances across every single feature dimension simultaneously. This creates boundaries that are mathematically sound but impossible for a human to visualize or describe.

Samuel Kaski and Teuvo Kohonen observed that even powerful visualization tools like Self-Organizing Maps (SOMs) fail to reveal which specific variables are responsible for the observed boundaries. This "interpretability gap" prevents domain experts from verifying whether a cluster represents a meaningful discovery or mere statistical noise.

Explainable by Design: Clustering with Decision Trees

The most prominent solution is to define clusters using axis-aligned decision trees. Instead of centroids, the algorithm uses a small number of "if-then" splits on individual features to partition the data.

Michal Moshkovitz and Sanjoy Dasgupta introduced the Iterative Mistake Minimization (IMM) algorithm, which was among the first to provide theoretical guarantees for this approach. By using a tree structure, the "explanation" for a cluster is simply the path from the root to the leaf. For example, a cluster might be defined as "Age > 40 AND Income < $50k," a description that is immediately actionable for a human analyst.

To make these trees even more efficient, Buddhima Gamlath and Ola Svensson developed "oblivious" algorithms that look only at cluster centers rather than the entire dataset, achieving high computational efficiency without sacrificing the simplicity of the resulting rules.

The Price of Explainability

Forcing clusters to follow simple, axis-aligned rules often comes at a cost to mathematical precision. This trade-off is known as the Price of Explainability (PoE).

Eduardo Sany Laber has extensively quantified this cost, showing that while explainable trees are slightly less "accurate" than unconstrained $k$-means, the loss is often negligible in practical scenarios. Lunjia Hu and Moses Charikar further improved these bounds, demonstrating that in low-dimensional spaces, we can achieve near-optimal clustering performance while maintaining total transparency.

Alternative Geometries: Polyhedra and Rules

While decision trees are popular, they are limited to axis-aligned splits. Other researchers explore more flexible geometric descriptions. Connor Lawless and Oktay Günlük proposed using polyhedral descriptions—wrapping clusters in shapes formed by the intersection of a few linear boundaries. This approach often provides higher descriptive accuracy than trees while remaining simple enough for human experts to validate.

In the realm of probabilistic modeling, Matthias Wilms and Jilles Vreeken developed Explainable Mixture Models (XMM). These models pair each component of a mixture with a differentiable rule, revealing the specific conditions that cause sub-populations to emerge in complex data.

Algorithm-Agnostic Explanations

Sometimes, users already have a preferred clustering algorithm and simply need to understand its output. Charles A. Ellis and colleagues introduced tools like G2PC and L2PC, which are "algorithm-agnostic." These methods perturb or shuffle specific features to see how often data points switch clusters, effectively identifying which variables drive the assignments regardless of whether the underlying model is $k$-means, spectral clustering, or a neural network.

Similarly, the E2L (Entity to Labels) framework by Alfio Ferrara and Stefano Montanelli repurposes techniques like LIME to identify "discriminative terms" for clusters, turning raw data groupings into human-readable labels.

Why It Works: Building Intuition

The core intuition behind this research is that structure should imply meaning. In a standard black-box model, structure is defined by proximity in a high-dimensional latent space that humans cannot perceive. Explainable clustering forces the model to find a structure that aligns with human cognitive constraints—specifically, our preference for small sets of discrete rules and low-dimensional boundaries.

By constraining the "language" the algorithm uses to define its groups (e.g., using a decision tree or a simple polyhedron), we bridge the gap between statistical discovery and human knowledge, transforming unsupervised learning from a labeling task into a tool for scientific and organizational insight.

Papers in this version

Showing 60 of 65 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_explainable_clustering_for_interpretable_ccbe99e3,
  title = {Explainable Clustering for Interpretable Unsupervised Learning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/explainable-clustering-for-interpretable-unsupervised-learning/rdr_ccbe99e3d37d400dafad5d3feecf1e9c},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_explainable_clustering_for_interpretable_ccbe99e3_v1,
  title = {Explainable Clustering for Interpretable Unsupervised Learning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/explainable-clustering-for-interpretable-unsupervised-learning/rdr_ccbe99e3d37d400dafad5d3feecf1e9c/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}