6 Clustering

Clustering and unlabeled data

In supervised learning, examples include a target label or value that a model learns from. In unsupervised learning, examples have no target labels, so a method looks for structure in the features themselves. Clustering is a common unsupervised task: it groups examples that are similar according to a chosen similarity measure.

For example, customer data might include purchase frequency, average order value, and product preferences without identifying customer types. Clustering can group customers with similar feature patterns; the groups are discovered from the data rather than supplied as correct answers.

Goals and choices

A clustering method assigns examples to groups, often called clusters. Its goal is to make examples within a cluster more alike—according to the method’s assumptions and similarity measure—than examples in different clusters. Depending on the method, it may also identify examples that do not fit well into any group.

Clustering can help people explore data for possible patterns or subgroups, summarize many examples with a smaller number of group profiles, or use group membership as one input to further analysis or decisions.

The resulting groups depend on which features are included, how those features are represented or scaled, what counts as similar, and which algorithm is used. Different choices can produce different groupings. Some methods require the number of clusters to be chosen in advance; others find groups according to different criteria.

Interpreting the groups

A cluster number or ID is only a label assigned by the algorithm. Cluster 0 is not inherently more important or meaningful than cluster 1. To interpret the groups, examine the original features of their members and describe what distinguishes each group.

For example, customer groups might be summarized as frequent purchases with lower average order value, infrequent purchases with higher average order value, or few purchases and low engagement. Such summaries suggest possible descriptions, but they do not prove that the groups are fixed or naturally occurring customer types.

A cluster may reflect an unhelpful feature, a data-preparation choice, or a pattern specific to the dataset. Consider whether the groups are understandable and useful for the intended purpose, and whether they remain reasonably consistent when appropriate choices or samples change.

Assessing results and limits

Because clustering does not usually have known correct labels, its results cannot be judged in exactly the same way as predictions with labeled answers. Internal measures and visualizations can help assess how separated or compact groups are, but they do not establish that the clusters are meaningful in context.

Clustering is useful for exploring and summarizing data, but its results depend on the features, similarity measure, and algorithm. Interpret groups by examining their feature patterns and usefulness rather than treating their IDs as pre-existing categories or ground truth.