Introduction

GridSearchCV is designed around repeated train-and-score splits, which is why its name ends with CV. In unsupervised learning, especially when you do not want cross-validation at all, the better question is usually not "how do I disable CV" but "what objective am I optimizing, and do I really need GridSearchCV for it?"

Why the usual GridSearchCV workflow is awkward here

In supervised learning, cross-validation is natural because labels define what it means to generalize. In unsupervised learning, there may be no ground-truth labels, so you often use an internal score such as silhouette score, inertia, or a domain-specific objective.

GridSearchCV still expects split logic. There is no clean built-in "no cross-validation" mode where it just fits once per parameter set with zero splitting.

That is why many unsupervised tuning workflows are clearer with ParameterGrid and a manual loop.

The simplest honest approach: manual grid search

Here is a direct grid search for KMeans using silhouette score.

1from sklearn.cluster import KMeans
2from sklearn.metrics import silhouette_score
3from sklearn.model_selection import ParameterGrid
4from sklearn.datasets import make_blobs
5
6X, _ = make_blobs(n_samples=300, centers=4, random_state=0)
7
8best_score = -1.0
9best_params = None
10
11for params in ParameterGrid({"n_clusters": [2, 3, 4, 5], "n_init": [10, 20]}):
12    model = KMeans(random_state=0, **params)
13    labels = model.fit_predict(X)
14    score = silhouette_score(X, labels)
15
16    if score > best_score:
17        best_score = score
18        best_params = params
19
20print(best_params, best_score)

This does exactly what many people mean by "GridSearchCV without cross-validation": it tries parameter combinations and scores each one on the available data.

If you insist on GridSearchCV

You can force GridSearchCV to use a custom split iterator that yields a single train-test split. That still is a split-based hack, not a true no-CV mode.

1import numpy as np
2from sklearn.cluster import KMeans
3from sklearn.metrics import silhouette_score
4from sklearn.model_selection import GridSearchCV
5from sklearn.datasets import make_blobs
6
7X, _ = make_blobs(n_samples=300, centers=4, random_state=0)
8
9
10def single_split(X, y=None, groups=None):
11    idx = np.arange(len(X))
12    yield idx, idx
13
14
15def silhouette_scorer(estimator, X, y=None):
16    labels = estimator.fit_predict(X)
17    return silhouette_score(X, labels)
18
19search = GridSearchCV(
20    estimator=KMeans(random_state=0),
21    param_grid={"n_clusters": [2, 3, 4, 5]},
22    scoring=silhouette_scorer,
23    cv=single_split(X),
24)
25
26search.fit(X)
27print(search.best_params_)

This works, but it is usually more confusing than the manual loop because training and scoring occur on the same data.

Choosing a scoring function

The hardest part of unsupervised tuning is not the grid. It is the score.

For clustering, common choices include:

• silhouette score
• Calinski-Harabasz score
• Davies-Bouldin score
• domain-specific business metrics

Each metric rewards different structure. A parameter set that minimizes inertia is not automatically the one that gives the most useful clusters.

Why cross-validation may still matter conceptually

Even when labels are absent, stability still matters. A clustering result that changes dramatically with small perturbations or random seeds may not be trustworthy.

That is why many practitioners evaluate multiple seeds, bootstrap samples, or downstream task performance rather than relying on a single score from a single full-dataset fit.

So "no cross-validation" can be operationally convenient, but it should not become a substitute for thinking about robustness.

Common Pitfalls

A common mistake is trying cv=1 and expecting GridSearchCV to become a no-CV tuner. That is not the right mental model, and it is not the clean solution.

Another issue is using a score that is incompatible with the estimator output. Some unsupervised metrics require predicted labels, some require distances, and some optimize in the opposite direction.

It is also easy to overinterpret internal metrics. A mathematically tidy cluster score does not guarantee clusters that are useful for the real business or scientific problem.

Summary

• 'GridSearchCV is fundamentally a split-based tool, not a pure no-CV tuner.'
• For unsupervised learning without cross-validation, a manual ParameterGrid loop is often the clearest solution.
• If needed, GridSearchCV can be coerced into a one-split setup, but that is a workaround.
• The quality of the tuning outcome depends heavily on the scoring metric you choose.
• In unsupervised settings, think about stability and usefulness, not just the best internal score.