6 Clustering

What clustering does

Clustering is an unsupervised learning task: an algorithm groups examples according to similarities in their features, without being given the correct group labels in advance. In classification, by contrast, a model learns to assign examples to predefined labels.

Clustering may reveal useful structure, but it does not automatically explain what each group means. People interpret the resulting groups in context.

Examples and applications

A store could represent each customer using features such as visit frequency and average purchase amount. A clustering algorithm might group customers with similar shopping patterns; the business could then inspect the groups and decide whether they correspond to occasional shoppers, regular customers, or another useful pattern.

Clustering can also support document and search-result grouping, image analysis, and exploratory analysis. For example, grouping pixels or regions with similar properties can support image segmentation. In scientific, health, or operational data, clustering may help identify patterns, unusual examples, or potential subgroups.

Text documents can be converted into feature vectors and clustered to discover groups of related topics. Such groups can help people explore a collection, but they should not be treated as definitive categories without review.

Features and similarity

A clustering method uses a similarity measure, or equivalently a distance measure, to decide which examples are alike. The appropriate measure depends on the features and the task.

Features with larger numerical scales can dominate distance calculations. Data often needs to be scaled or transformed before clustering. Because algorithms define clusters differently, they can produce different groupings from the same data.

How clustering methods differ

K-means assigns examples to the nearest of a chosen number of centers, then updates the centers and repeats. It is efficient for compact, roughly round clusters, but requires choosing the number of clusters and can be affected by outliers or its starting centers.

Hierarchical clustering builds a hierarchy of groups. In an agglomerative approach, each example starts alone and the closest groups are repeatedly merged. The resulting tree can be cut at different levels to obtain different numbers of clusters.

DBSCAN groups examples in dense regions and can mark isolated examples as noise. It can find irregularly shaped clusters, but its results depend on settings such as neighborhood radius and minimum neighborhood size.

Gaussian mixture models represent data as a mixture of probability distributions. Rather than assigning each example only to one group, they can estimate its probability of belonging to each cluster.

These methods rely on different assumptions and suit different data shapes. No single method is best for every dataset.

Interpreting and evaluating clusters

A cluster is a pattern produced under a particular choice of features, distance measure, and algorithm; it is not necessarily a naturally existing or meaningful category. Consider whether the groups are useful for the intended task, whether results change substantially with different settings, and whether the examples grouped together make sense.

If known labels are available, they can help compare clusters with those labels. Without labels, evaluation is less direct and may use measures such as the silhouette coefficient alongside practical review.