Learn AI /How Machines Learn /Clustering
Clustering
- Intermediate
- How Machines Learn
- Technical readers
- Programme & data teams
Also called: cluster analysis.
Back to the milestone you came from
In five seconds
What it is
A method that groups similar examples together so patterns in data can be found without pre-written categories.
In plain words
Clustering groups similar items together when you do not already have named categories.
An everyday example
At a busy coach station, a worker sees a pile of untagged bags. They start placing bags near each other by visible similarities: small hard cases together, large hiking packs together, and medium fabric bags together. After a while, groups appear even though nobody provided official bag categories. A supervisor then looks at the groups and says which ones seem useful for storage, loading or extra inspection.
Real clustering does not look at objects the way a person does, and it does not automatically know which similarities matter operationally. It works on numerical representations of features, and different feature choices can produce very different groups. Also, real data may not fall into neat piles at all; some points sit between groups or form uneven shapes.
Remember thisClustering proposes groups from similarity in selected features. Change the features, scaling, distance measure or parameters, and the groups may change.
Understand it
Understand it
Why it matters
Clustering can help people notice structure in messy data when no categories have been defined yet. That makes it useful for exploration, segmentation, anomaly review and early pattern detection. For a non-specialist, the key decision point is this: a cluster can suggest where to look, but it should not be treated as proof of a real-world group, risk or cause. Results can change if you choose different features, scales or algorithms. Good use of clustering means asking whether the grouping is stable, useful for the task and safe to act on, especially when people or services may be affected.
How clustering forms groups from unlabeled data
After viewing this, the reader should be able to say that clustering starts with unlabeled examples, compares them by features, and outputs suggested groups whose meaning must be interpreted.
01
1. Start with examples
Each point is one example. Its position comes from two features, shown on the horizontal and vertical axes.
Draw a left panel with 12 to 20 grey dots scattered across an x-y plane labelled “feature 1” and “feature 2”. Add note: “no labels provided”.
02
2. Compare similarity
The method measures which points are close to each other in feature space. Closeness depends on the chosen features and distance rule.
Draw a middle panel with faint lines or shaded neighbourhoods showing nearby points influencing each other. Keep points grey. Add caption: “group by similarity”.
03
3. Form clusters
Points that are more similar to one another than to others are assigned to groups. Some points may be near a boundary or left alone, depending on the method.
Draw a right panel with the same points now coloured into 3 groups, plus one boundary point shown with a dashed outline. Surround groups with soft contours.
04
4. Interpret carefully
The coloured groups are proposed structure, not ground truth. A person checks whether the grouping is useful for the real problem.
Add small text labels beside the contours such as “possible pattern A”, “possible pattern B”, “possible pattern C”, with a footer note: “meaning comes from interpretation and validation”.
Show scattered data points on a two-dimensional feature space, then show points gathering into colour-coded groups, then show labels added by a human as interpretations. Read left to right: examples, similarity grouping, human interpretation.
Hovering or tapping a point highlights its nearest neighbours in step 2 and its assigned group in step 3; a toggle switches between different feature scalings to show that clusters can change.
Describe this without the diagram
The diagram has four left-to-right stages. First, it shows many unlabeled data points on a graph with two features. Second, it shows that the method compares nearby points based on similarity in those feature values. Third, it colours the points into three groups and marks one point near a boundary to show that membership can be uncertain or method-dependent. Fourth, it adds text saying the groups are only proposed patterns and require human interpretation. The main message is that clustering finds structure from feature similarity, but the result depends on representation and does not create proven real-world categories by itself.
With reduced motion: A static four-panel diagram is shown instead of animated movement. The first panel shows grey unlabeled points, the second panel shows neighbourhood shading, the third panel shows coloured groups with one dashed boundary point, and the fourth panel adds interpretation notes.
Clustering is a way to organise data by similarity. You start with examples and their features, meaning measurable properties such as size, location, frequency or word counts. Unlike classification, the examples do not come with known category labels in advance.
What goes in is a set of examples, each described by features. The method compares examples and looks for ones that are close to each other in that feature space, which means their feature values are similar by some mathematical measure. Different clustering methods define “close” in different ways.
What changes inside depends on the method, but the general aim is similar: place examples into groups so items in the same group are more alike than items in different groups. Some methods require you to choose the number of groups first. Others try to find dense regions or hierarchical structure.
What comes out is not a fact about the world. It is a proposed grouping based on the features used, the scaling of those features and the algorithm chosen. That is why clusters are useful for exploration, but they still need human interpretation and checking against the real task.
How “Sorting mixed luggage at a busy coach station” maps onto the real thing
| In the story | In the AI system |
|---|---|
| untagged bags | examples without labels |
| size, shape and material of each bag | features used to compare examples |
| placing similar bags near each other | grouping examples by similarity |
| supervisor interpreting the piles | human review of what the clusters mean |
A worked example
A facilities management officer at the Bogotá duty station reviews readings from nine chilled-water pumps before updating the maintenance workplan. Each record contains average vibration and motor temperature.
Starting from: The records are P1=(2.1,48), P2=(2.4,50), P3=(2.0,47), P4=(7.8,82), P5=(8.2,79), P6=(7.5,85), P7=(4.9,61), P8=(5.1,60) and P9=(5.3,63).
- The officer represents each pump reading as two features: vibration and temperature.
- Because the features use different units and scales, each is standardised before clustering.
- A k-means procedure is run with three clusters. It repeatedly assigns each record to the nearest centre and recalculates the centres until the assignments stabilise.
- The resulting groups contain P1 to P3, P7 to P9, and P4 to P6. The officer reviews the high-reading group first, alongside maintenance history and an inspection.
Result: The method suggests three patterns of operating readings for the maintenance workplan. It does not establish that any pump is faulty, and a different feature set, scaling choice or number of clusters could produce different groups.
In a United Nations workplace
At a duty station working in three languages, a duty officer must screen a long list of applications against a rostered profile by close of business. Using an approved internal tool, the officer clusters applications by selected job-related features, such as functions performed, type of experience and stated language skills, to identify recurring patterns before reviewing each file against the rostered criteria.
Clustering can reveal groups of applications with similar profiles without assigning pre-written categories. It may help organise the file, but it does not show whether an applicant meets the rostered profile or should be shortlisted.
Before you try thisApplications contain personal and HR-sensitive data. They must not be uploaded to any tool not approved for that material. Clusters must not determine eligibility, ranking or shortlisting: they reflect selected features and may reproduce bias, miss equivalence across languages, or change when scaling, distance measures or parameters change. The rostered criteria and human review remain controlling.
Where you may hear it
- A vendor demonstration to an ICT governance board
- An evaluation office reviewing an analysis plan
- An inter-agency working group examining a needs assessment
- A programme section discussing patterns in monitoring reports
Avoid confusion
Avoid confusion
Often confused with
Classification
A machine learning method that assigns one or more predefined category labels to an input based on learned patterns.
Prediction
The result a model produces from input, such as a class, number, ranking or next token.
Topic modelling
A method that scans many documents and groups frequently co-occurring words into recurring themes, helping people see what a collection is mostly about.
Common misconceptions
Common misconception: Clustering discovers the one true set of categories hidden in the data.
Correction: Usually there is no single correct grouping. Different algorithms, feature choices and scaling decisions can produce different clusters, and several may be useful for different purposes. The output is a model-based grouping, not automatic ground truth.
Common misconception: Clustering and classification are basically the same thing.
Correction: They both assign examples to groups, but classification learns from examples with known labels, while clustering starts without labels and tries to find structure from similarity alone. In classification, the target categories are given beforehand; in clustering, they are not.
Common misconception: If points are in the same cluster, they must share the same cause or meaning.
Correction: A cluster only shows similarity in the chosen features. Two examples may look similar numerically for different real-world reasons. Human review and domain knowledge are needed before drawing operational conclusions.
Quick check
Which statement best describes clustering?
Clustering is about grouping unlabeled examples by similarity. Option 1 describes classification, not clustering. Option 3 is too strong because many methods need parameter choices and may produce several plausible groupings. Option 4 describes a narrow anomaly-detection idea, not clustering in general.
Go deeper
Go deeper
The simple version
Clustering groups similar items together when you do not already have named categories.
The technical version
Clustering is an unsupervised learning task that partitions or organises examples in a feature space according to a similarity or distance criterion. Depending on the algorithm, it may minimise within-cluster variance, identify dense regions separated by sparse regions, or build a hierarchy of nested groups. Common approaches include k-means, hierarchical clustering and density-based methods. The result is sensitive to feature engineering, scaling, distance metric, parameter choices and assumptions about cluster shape, size and number. Because no labels are supplied, evaluation is often indirect: internal measures such as compactness and separation can help, but external usefulness depends on the downstream task and domain interpretation.
Reviewed external resources
EncyclopaediaClustering (opens in a new tab)Wikipedia
This page gives a broad overview of clustering methods and terminology, and it suits readers who want a general reference.
Intermediatelink checked 2026-07-31
CourseMachine Learning Crash Course (opens in a new tab)Google
This course resource suits learners who want practical machine learning context and a structured introduction at an accessible level.
Intermediatelink checked 2026-07-31
VideoStatQuest (opens in a new tab)StatQuest with Josh Starmer
This video source suits readers who prefer step-by-step spoken explanations with intuition before more formal detail.
Starterlink checked 2026-07-31
VideoVideo explanations: Clustering (opens in a new tab)YouTube search
This is a YouTube search results page, so it helps readers browse multiple video explanations and choose one that matches their level.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Clustering (opens in a new tab)Reddit search
This is a Reddit search results page, so it suits readers who want to see community questions, informal explanations and common points of confusion.
search results, not one pagelink checked 2026-07-31
Where this came from
Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.
The one thing to remember
Clustering proposes groups from similarity in selected features. Change the features, scaling, distance measure or parameters, and the groups may change.