Using Decision Trees to Explain Clustering Results#

Stage 7 · 🌳 Decision Trees · Lesson 49 of 56 · advanced

◀ Previous · Cluster Profiling Using Decision Trees · Next · Assessing the Quality of Prediction Models ▶ · ↑ Section

Important

✨ AI-generated content. This page was written with the assistance of an AI language model and is provided as a learning aid. Despite careful review, it may still contain mistakes, omissions, or out-of-date information. Whether you are new to the topic, a team lead, or a senior practitioner, treat it as a starting point rather than an authoritative reference: read it critically and independently verify anything you act on (code, commands, figures, and factual claims) against official documentation and primary sources before relying on it.

Explaining a black box#

The last lesson profiled individual clusters; this one generalises the idea into one of machine learning’s most useful tools. Clustering algorithms — and many models besides — are opaque: k-means or a neural network produces outputs, but not an explanation of its logic. A surrogate model is a deliberately simple, interpretable model trained to mimic the opaque one, so that its transparent logic can stand in for the black box’s hidden one.

The surrogate idea#

The recipe is general. Take whatever the opaque method produced — here, the cluster label for each observation — and treat those outputs as the target. Then fit an interpretable model, most naturally a decision tree, to reproduce them from the features. The tree learns the rules that best explain the clustering as a whole, and because it is a tree, those rules can be read: a compact, global description of how the clustering carved up the data. This is a global surrogate — it approximates the whole model’s behaviour, not just one prediction — and it is model-agnostic, needing no knowledge of how the clustering worked inside.

Fidelity#

The obvious question is: can you trust the surrogate? A tree that mimics the clustering only loosely would give a misleading explanation. The measure of trust is fidelity — how faithfully the surrogate reproduces the original’s outputs, typically the surrogate’s accuracy at predicting the black box’s labels. High fidelity means the tree’s rules genuinely reflect the clustering; low fidelity means the explanation is unreliable. There is a real tension here: a deeper tree fits the clustering more faithfully but is harder to read, so surrogates trade fidelity against interpretability, and a shallow, high-fidelity tree is the happy case.

A general pattern#

This surrogate trick is far bigger than clustering. The same move — fit an interpretable model to a complex model’s outputs, then check its fidelity — explains black-box classifiers, regressors and ensembles too, and it complements the per-prediction attributions of Shapley values from Stage 5. It closes the trees stage on a fitting note: the decision tree, prized for its transparency, becomes a lens for seeing into models that have none. With the workflow’s models built and understood, the final stage asks the essential question — how good are they?

Hint