Must a centroid be a data point?

No. The arithmetic mean often lies between observations.

Is a centroid enough to describe a cluster?

No. Size, spread, shape, and representative observations supply information the center omits.

AI Glossary / Mathematics and optimization / Intermediate

Cluster centroid

Q: Why does k-means use the mean?

For fixed assignments, the mean minimizes the sum of squared Euclidean distances within each nonempty cluster.

Definition

A cluster centroid is a representative center used to summarize a cluster of data points. In ordinary Euclidean k-means, it is the coordinate-wise arithmetic mean of the points assigned to that cluster. Other clustering methods and geometries can use different definitions of a representative center, so the convention should be stated.

What it means

A centroid condenses many numerical feature vectors into one vector. For example, a group described by two measurements has a centroid with the average value of each measurement. The resulting point need not be an actual observation.

This entry emphasizes the familiar arithmetic-mean centroid used in k-means. The supplied glossary also allows broader choices, such as a geometric median or centers defined in non-Euclidean geometry. These choices do not share every property of the mean. A centroid is therefore best understood together with the representation, distance, and clustering objective that give it meaning.

How it works

For n equally weighted vectors x_1 through x_n, the mean centroid is their sum divided by n. Compute this separately for every coordinate. If observations have nonnegative weights with a positive total, the weighted centroid divides the weighted sum of vectors by the total weight.

The mean minimizes the sum of squared Euclidean distances to a fixed collection of points. This explains its role in k-means: an assignment step puts points with their nearest current center, and an update step recalculates each center from its assigned points.

That property does not imply the algorithm finds the globally best clustering. Initialization, the selected number of clusters, feature scaling, and local solutions still matter. Nor does a center describe the whole shape or spread of a cluster. Two groups can have the same mean while differing substantially in dispersion or internal structure.

Three practical examples

1. Customer analysis: summarize the numerical feature profile of a segment, while checking actual members before describing it as a typical customer.

2. Image analysis: use centroids of color vectors to represent groups in a simplified color palette, under a specified color representation and distance.

3. Embedding exploration: summarize a group of document vectors and examine nearby real documents to understand what the numerical center represents.

Think of it this way

Think of balancing equal weights placed at the data points. Their balance point is the arithmetic-mean centroid. Larger observation weights pull the balance point more strongly toward their locations.

The analogy explains the mean in Euclidean coordinates. It does not make the center a real member of the group, establish that the group is meaningful, or extend unchanged to every distance or geometry used in clustering.

Worked example

Take three two-dimensional points: A=(1,1), B=(3,1), and C=(2,4). Their centroid is ((1+3+2)/3,(1+1+4)/3)=(2,2).

The squared distances to this center are 2 for A, 2 for B, and 4 for C, giving a total of 8. Compare the candidate center (2,1): its squared distances are 1, 1, and 9, totaling 11. The mean provides the smaller value, consistent with its least-squares property.

Now add D=(10,2). The new centroid is ((1+3+2+10)/4,(1+1+4+2)/4)=(4,2). A single distant point moves the horizontal coordinate from 2 to 4, illustrating the mean's sensitivity to extreme observations.

For a separate weighted example using only A, B, and C, give C weight 2 and the others weight 1. The weighted centroid is ((1+3+2×2)/4,(1+1+2×4)/4)=(2,2.5). These examples assume meaningful numerical coordinates; averaging category codes such as department identifiers would not automatically produce a meaningful center.

Illustrative case study

Fictional case: a service company clusters accounts using monthly request volume and average resolution time. Its analyst presents each centroid as a real account profile.

A reviewer notices that one center lies between two distinct kinds of accounts and corresponds to no actual customer. The team keeps the centroid as a numerical summary but adds representative observed accounts and measures of spread. It also documents how the features were scaled, because changing the units can change which accounts are grouped together. The revised explanation is more faithful to the data than the center alone.

Common misconceptions and mistakes

A centroid is not necessarily a medoid. A medoid is selected from the observed points according to a distance-based objective; an arithmetic-mean centroid can lie between them.

Do not assume every clustering algorithm requires a centroid. Do not overlook outliers or incompatible feature scales. A centroid of a ring-shaped group can sit in an empty central region, so location alone does not describe density. Finally, updating an empty cluster has no ordinary arithmetic mean and needs a deliberate algorithmic policy.

Business use

Centroids help teams describe segments, build compact representations, and compare changes in group profiles. Their usefulness depends on transparent feature definitions and an interpretation that respects the underlying data.

Report cluster size and dispersion alongside the center, and inspect actual records before assigning a business label. When comparing centroids over time, keep preprocessing consistent or explain changes. A moved center may reflect a different population mix, altered measurement, or changed assignments rather than a uniform change among all members.

Check your understanding

The points (0,0) and (4,2) have equal weight. What is their centroid, and is it necessarily an observed point?

Answer: Their centroid is (2,1). It is not one of the two observations, illustrating why a center should not automatically be presented as an actual member.

Related terms

k-means
Medoid
Geometric median
Euclidean distance
Clustering
cluster mean
feature scaling
within-cluster sum of squares

Learn next

compare means, medoids, and geometric medians on data with outliers. Then examine how feature scaling changes nearest-center assignments and why reporting within-cluster variation is essential for interpretation.

Sources

Aalto Dictionary of Machine Learning, cluster centroid, clustering, and Lloyd's algorithm entries: https://aaltodictionaryofml.github.io/ADictML_Main.pdf . Relevant definitions and the broader center convention were read. All coordinate calculations and the fictional interpretation example are original.

Previous: Cluster assumption Next: Cluster hypothesis