Introduction

K-means requires you to choose k, the number of clusters, before fitting the model. So the honest answer is that you cannot run true k-means without some value of k; what you can do is estimate a reasonable k from the data or switch to a different clustering algorithm that does not need that parameter.

What K-means Assumes

K-means is built around a simple objective: assign each point to one of k centroids so that within-cluster variance is minimized. The value of k is part of the problem definition, not an optional detail.

That is why the question usually becomes one of these instead:

• how do I choose k sensibly
• how do I compare several candidate values of k
• should I use a different clustering method altogether

Common Ways to Choose k

Several heuristics are used in practice.

• elbow method
• silhouette score
• gap statistic
• domain knowledge or downstream business constraints

None of these is perfect. They are model-selection tools, not mathematical proof that one cluster count is uniquely correct.

A Practical Example with Silhouette Score

A common workflow is to fit k-means for a range of candidate values and pick the one with the best validation signal.

1import numpy as np
2from sklearn.cluster import KMeans
3from sklearn.metrics import silhouette_score
4
5X = np.array([
6    [1.0, 1.0], [1.2, 0.9], [0.8, 1.1],
7    [5.0, 5.0], [5.2, 4.8], [4.9, 5.1],
8    [9.0, 1.0], [9.2, 0.8], [8.8, 1.1],
9])
10
11best_k = None
12best_score = -1.0
13
14for k in range(2, 6):
15    model = KMeans(n_clusters=k, n_init=10, random_state=42)
16    labels = model.fit_predict(X)
17    score = silhouette_score(X, labels)
18    print(k, score)
19
20    if score > best_score:
21        best_score = score
22        best_k = k
23
24print("chosen k:", best_k)

This does not eliminate uncertainty, but it gives you a repeatable way to compare candidate values.

When K-means Is the Wrong Tool

Sometimes the real problem is not choosing k. It is that k-means assumes compact roughly spherical clusters and uses centroid distance as the organizing principle.

If your data has:

• irregular cluster shapes
• varying density
• lots of noise or outliers
• no obvious centroid structure

then algorithms such as DBSCAN or HDBSCAN may be more appropriate because they do not require a fixed cluster count up front in the same way.

That can be a better answer than forcing k-means onto data it does not model well.

Use Domain Knowledge When It Exists

Cluster validation metrics are useful, but domain knowledge often matters more.

For example, if the business question is "segment customers into five actionable campaign groups," then the practical answer might still be k = 5 even if another metric slightly prefers k = 4 or k = 6.

Clustering is often exploratory, not purely mathematical. The chosen k has to serve the use case, not only the score.

Stability Matters Too

A good check is to run k-means several times with different random initializations and see whether the structure is stable. If small changes in initialization lead to very different cluster assignments, that is a warning that the data may not support a clean k-means partition.

Modern scikit-learn already supports multiple initializations with n_init, and you should use that instead of trusting a single run.

Common Pitfalls

Expecting k-means itself to infer k automatically is the most common misunderstanding. That is outside the algorithm's contract.

Using the elbow method as though it always has a clear elbow also causes overconfidence. Many datasets produce ambiguous curves.

Treating the highest silhouette score as absolute truth is another mistake. A numerically good score may still correspond to a useless business segmentation.

Finally, if the data shape is incompatible with centroid-based clustering, no amount of tuning k will make k-means the right model.

Summary

• k-means requires k; it cannot run without some chosen cluster count
• in practice, you estimate k using heuristics such as the elbow method or silhouette score
• domain knowledge often matters as much as any clustering metric
• if the data has irregular shapes or strong noise, consider algorithms such as DBSCAN instead of forcing k-means
• the real goal is not only picking a number, but choosing a clustering model that matches the structure of the data