Does it find the true number of groups?

It produces a count under its assumptions and selection rules; a unique true count may not exist for the task.

Can it return no clusters?

Yes. Some density-based methods can label every observation as noise under restrictive settings.

Is it the same as classification?

No. Classification typically learns from predefined labels, while clustering explores structure without those target labels.

AI Glossary / Data science and analytics / Advanced

Automatic clustering algorithms

Definition

Automatic clustering algorithms are methods that form groups in unlabeled data while determining the number of groups through a procedure rather than requiring a fixed cluster count in advance. Some infer groups from density or similarity; others compare candidate clusterings using a selection criterion. Automatic selection does not guarantee a uniquely correct grouping or remove the need for modeling choices.

What it means

Clustering asks which observations resemble one another under a chosen representation. Many familiar methods require the analyst to choose a number of groups first. Automatic approaches move some of that decision into the algorithm or a surrounding selection process.
The word automatic can therefore describe different capabilities. One method may choose the cluster count but still require a distance threshold. Another may search across counts and select the best according to a score. Neither determines the analysis purpose. A useful grouping of support requests by technical cause may differ from a useful grouping by the team responsible for answering them.

How it works

Prepare features and define similarity or distance. Choose a method and specify the settings that control its notion of a group. For example, density-based methods such as DBSCAN can produce a cluster count from neighborhood and density settings, with some observations labeled as noise. Affinity propagation chooses representative exemplars, but its preference parameter influences how many emerge.
Alternatively, run a clustering method across candidate counts and apply a selection rule. The selected result depends on that rule, the candidate range and the data representation.
Inspect the resulting groups, unassigned observations and sensitivity to reasonable setting changes. Compare results across resampled or later data where appropriate. Stability is useful evidence, but does not prove the groups match the intended business concept.

Three practical examples

A support team groups incoming issue descriptions without fixing the number of topics beforehand. Reviewers inspect the proposed groups and keep unusual tickets available for investigation rather than treating noise as disposable data.
A publisher explores similarities among glossary articles to identify possible navigation categories. It checks that the suggested group count produces understandable categories and does not simply reflect repeated formatting or citation patterns.
An operations analyst groups equipment behavior using measured features. Before interpreting the groups as operating modes, the analyst checks whether units, scaling or missing observations drove the separation.

Think of it this way

Imagine asking someone to sort objects without telling them how many boxes to use. They still need a rule for deciding what belongs together: color, shape, purpose or something else. An automatic cluster count is the number of boxes produced under that rule. It is not proof that every other reasonable sorting rule would produce the same answer.

Worked example

Consider one-dimensional points at 0, 0.1, 0.2, 5, 5.1, 5.2 and 10. Apply DBSCAN with Euclidean distance, neighborhood radius 0.25 and a minimum of three points, counting the point itself, to qualify as a core point.
Each point in the first group has the other two within radius 0.25. The same holds for the three points near 5. These form two clusters. The point at 10 has only itself in its neighborhood, so it remains noise. The algorithm returns two clusters without receiving a requested cluster count of two.
Now reduce the radius to 0.05 while keeping the minimum at three. No point has enough neighbors, so all seven are noise and there are zero clusters. This invented example shows that avoiding a fixed cluster-count input does not make the result parameter-free. Noise describes the algorithm's treatment of a point under these settings, not proof that the observation is erroneous.

Illustrative case study

In a fictional project, Maple Support wants to organize a growing archive of help requests. The team compares a fixed-count baseline with a method that determines groups from local density. Its first automatic result contains several coherent groups and a substantial collection of unassigned tickets.
Reviewers find that some unassigned tickets describe rare but important issues. They also discover a group dominated by a standard email signature rather than a shared problem. The team improves text preparation, repeats the analysis and reviews representative tickets from each proposed group.
Maple keeps the unassigned queue and measures whether the grouping helps staff retrieve relevant previous cases. It documents the chosen settings and data period. This fictional project illustrates review; the selected count is not an objective inventory of all support topics.

Common misconceptions and mistakes

Automatic does not mean parameter-free, assumption-free or robust to every outlier. Different methods define clusters differently and can legitimately produce different results. Ordinary k-means does not choose its own cluster count unless another procedure is added around it.
Selecting the smallest within-cluster error without a complexity control can favor excessive fragmentation. A validation score also has its own preferences, so optimizing it is not identical to optimizing usefulness.
Cluster labels are identifiers, not rankings or permanent meanings. A later run's cluster 2 may not correspond to the earlier run's cluster 2. Compare membership and interpretation rather than matching label numbers alone.

Business use

Use automatic clustering when a fixed count would be arbitrary and exploratory grouping can support a defined decision. Record the features, distance measure, selection rule and settings.
Review group coherence, coverage, stability and the effort required to handle exceptions. Compare with a simple baseline and involve people who understand the records. If clusters will drive operational actions, validate those actions separately and plan for new observations. Some methods describe the fitted dataset without providing a straightforward rule for assigning future cases.

Check your understanding

Why did the example change from two clusters to none?

Answer: The smaller radius prevented any point from satisfying the core-point requirement.

Why inspect unassigned observations?

Answer: They may contain valid, rare or operationally important cases.

What remains a human decision?

Answer: The purpose, representation, acceptable assumptions and interpretation of the grouping.

Related terms

Unsupervised learning
Noise
Cluster analysis
DBSCAN
affinity propagation
cluster validation
silhouette score
hyperparameter
feature scaling

Learn next

Repeat a small clustering exercise across several distance thresholds and record both cluster counts and unassigned points. Then compare the numerical result with the grouping needed for a concrete decision.

Sources

Scikit-learn, Clustering user guide: method assumptions, parameters and evaluation concepts. https://scikit-learn.org/stable/modules/clustering.html Scikit-learn, affinity_propagation documentation: preference influences the number of exemplars. https://scikit-learn.org/stable/modules/generated/sklearn.cluster.affinity_propagation.html The numerical DBSCAN example and Maple Support case are original illustrations. No universal optimality or outlier-resistance claim is intended.

Previous: Automatic assembly sequencing Next: Automatic differentiation