Clustering is an unsupervised machine learning method used to group similar data points together. It helps find structure in data without predefined labels, so analysts and models can discover segments, patterns, clusters and unusual points.

In machine learning, clustering is used when we do not already know the correct categories. The algorithm looks at the data and tries to organize similar points into groups.

For example, clustering can group customers with similar behaviour, documents with similar topics, products with similar properties, images with similar visual features or transactions with similar patterns. The goal is not to predict a known label. The goal is to discover useful structure.

What clustering means

Clustering is a method for finding groups in data. A cluster is a group of data points that are more similar to each other than to points in other groups.

For example, if you analyse website visitors, clustering may reveal groups such as price-sensitive users, loyal repeat customers, one-time buyers, B2B visitors or inactive users. These groups may not be written in the dataset. The clustering algorithm tries to discover them from patterns in the data.

Clustering is called unsupervised because the model does not receive correct answers during training. There are no predefined labels such as „good customer“, „bad customer“ or „fraud“. The algorithm works from similarity, distance, density or structure.

A simple example of clustering

Imagine you have a basket of different fruits. Nobody tells you the names of the fruits. You only see their properties: size, colour, weight, shape and texture.

A clustering algorithm may group small red round fruits together, long yellow fruits together and large green fruits together. It does not know the words apple, banana or melon. It only sees similarity.

This is the basic idea of clustering. It finds groups based on patterns in the data, not based on human labels.

Why clustering is used

Clustering is used when data contains structure, but the structure is not labelled in advance.

It is useful for:

• customer segmentation – grouping customers by behaviour or value,
• document clustering – grouping texts by topic or meaning,
• image analysis – grouping visually similar images,
• product grouping – finding similar products or product families,
• fraud analysis – finding unusual groups of transactions,
• outlier detection – identifying points that do not belong to any normal group,
• exploratory data analysis – understanding structure before building models,
• recommendation systems – grouping similar users, items or content.

Clustering helps people see patterns that may be difficult to find manually.

Clustering and unsupervised learning

Clustering is one of the main types of unsupervised learning.

In supervised learning, the model learns from examples with known answers. For example, the training data may already say whether an email is spam or not spam.

In clustering, the model does not receive these answers. It only receives input data and tries to find structure.

This makes clustering useful when:

• labels do not exist,
• labels are expensive to create,
• the goal is exploration,
• the categories are unknown,
• the analyst wants to discover hidden groups.

How clustering works in simple terms

Different clustering algorithms work in different ways, but the general idea is similar. The algorithm needs a way to decide which data points are similar.

A simplified clustering process looks like this:

• Represent the data – each item is described using features or embeddings.
• Measure similarity – the algorithm compares points using distance or similarity.
• Form groups – similar points are placed together.
• Evaluate the result – the analyst checks whether the groups are meaningful.
• Interpret the clusters – each group is described using business, statistical or domain context.

The quality of clustering depends heavily on data representation. If the features are poor, the clusters will also be poor.

What a cluster is

A cluster is a group of similar data points. Similarity depends on the chosen features, distance metric and algorithm.

For example:

• customers may be similar because they buy the same products,
• documents may be similar because they discuss the same topic,
• images may be similar because they contain the same object,
• transactions may be similar because they have the same behavioural pattern,
• sensor readings may be similar because they represent the same machine state.

Clusters are not always real-world categories. Sometimes they are useful analytical groups. Sometimes they are artefacts created by the algorithm.

Similarity and distance

Clustering depends on similarity. In many algorithms, similarity is measured by distance.

If two points are close, they are treated as similar. If they are far apart, they are treated as different.

Common distance or similarity measures include:

• Euclidean distance – common for numerical data,
• Manhattan distance – useful when absolute differences matter,
• cosine similarity – often used for text embeddings,
• Jaccard similarity – useful for sets or binary data,
• correlation distance – useful when shape or trend matters more than scale.

The distance metric matters. A poor metric can create misleading clusters.

Clustering and feature quality

Clustering is sensitive to the features used.

If the features describe meaningful structure, clustering may reveal useful groups. If the features are noisy, irrelevant or badly scaled, clustering may create groups that do not mean much.

For example, customer clustering based on purchase recency, frequency and value may be useful. But clustering based on random tracking IDs, temporary campaign tags or inconsistent browser data may create meaningless groups.

Good clustering usually requires:

• clean data,
• relevant features,
• proper scaling,
• domain knowledge,
• validation of the result,
• careful interpretation.

Scaling before clustering

Scaling is often important before clustering because many algorithms depend on distance.

If one feature has values between 0 and 1 and another has values between 0 and 1 000 000, the large-scale feature may dominate the distance calculation.

For example, in customer clustering, annual revenue may dominate age, visit frequency or engagement score if features are not scaled.

Common scaling methods include:

• standard scaling – centres values around mean and standard deviation,
• min-max scaling – maps values to a fixed range,
• robust scaling – uses median and quartiles, useful when outliers exist,
• log transformation – useful for skewed variables such as revenue or traffic.

Scaling should be chosen based on the data and the algorithm.

K-means clustering

K-means is one of the best-known clustering algorithms. It divides data into a chosen number of clusters, called k.

The algorithm works by assigning each point to the nearest cluster centre, called a centroid. Then it updates the centroids and repeats the process until the clusters stabilise.

K-means is popular because it is simple, fast and scalable. It is often a good first method for clustering numerical data.

However, k-means has limitations. It works best when clusters are roughly round, similar in size and separated in a way that fits centroid-based grouping. It is also sensitive to outliers, scaling and the chosen number of clusters.

Choosing the number of clusters

Some clustering methods require the analyst to choose the number of clusters. K-means is the typical example.

Choosing the number of clusters is not always easy. Too few clusters may hide important differences. Too many clusters may create artificial groups.

Common methods include:

• elbow method – looks for a point where adding more clusters gives diminishing improvement,
• silhouette score – measures how well points fit within their assigned cluster compared with other clusters,
• domain knowledge – uses business or scientific understanding,
• stability testing – checks whether clusters remain similar across samples or settings,
• business usefulness – asks whether the clusters can be understood and acted on.

The best number of clusters is not always the one with the best mathematical score. It also needs to make sense for the task.

Hierarchical clustering

Hierarchical clustering creates a tree-like structure of clusters. This structure is often shown as a dendrogram.

There are two main approaches:

• agglomerative clustering – starts with each point as its own cluster and merges them step by step,
• divisive clustering – starts with one large cluster and splits it into smaller clusters.

Hierarchical clustering is useful when you want to see relationships at different levels. For example, documents may first split into broad topics and then into smaller subtopics.

It can be easier to interpret than some other methods, but it can become expensive on very large datasets.

DBSCAN

DBSCAN stands for Density-Based Spatial Clustering of Applications with Noise. It finds clusters as dense regions separated by areas of lower density.

Unlike k-means, DBSCAN does not require choosing the number of clusters in advance. It can also identify points as noise or outliers.

This makes DBSCAN useful when:

• clusters have irregular shapes,
• the number of clusters is unknown,
• outliers should be detected,
• density structure matters.

However, DBSCAN can struggle when clusters have very different densities. It also depends strongly on parameter choices such as epsilon and minimum samples.

Density-based clustering

Density-based clustering methods define clusters as dense regions of points. Points in sparse regions may be treated as noise.

This approach is useful when clusters are not round. For example, a cluster may form a curved shape, a ring or another irregular structure.

Density-based clustering is often better than k-means when:

• clusters are not spherical,
• outliers matter,
• the number of clusters is not known,
• dense regions are more meaningful than centroids.

DBSCAN and HDBSCAN are common examples of this family.

Gaussian mixture models

Gaussian mixture models, often shortened to GMMs, treat data as coming from a mixture of probability distributions.

Instead of assigning each point only to one cluster with complete certainty, a Gaussian mixture model can assign probabilities. A point may belong mostly to one cluster but partly to another.

This is useful when clusters overlap or when uncertainty matters.

For example, a customer may be 70 % similar to one segment and 30 % similar to another. This can be more realistic than forcing every customer into exactly one group.

Spectral clustering

Spectral clustering uses graph-based ideas. It builds a similarity graph between points and then finds groups based on the structure of that graph.

It can work well for complex cluster shapes where k-means may fail.

However, spectral clustering can be more computationally expensive and may require careful parameter choices. It is often useful for smaller or medium-sized datasets where relationships between points are important.

Clustering vs classification

Clustering and classification are different tasks.

Classification uses predefined labels. The model learns to assign new examples to known categories.

Clustering does not use predefined labels. It discovers groups from the data itself.

For example:

• Classification – predict whether an email is spam or not spam using labelled examples.
• Clustering – group emails by topic without knowing the topics in advance.

Classification is supervised learning. Clustering is unsupervised learning.

Clustering vs segmentation

Segmentation is often the business use of clustering.

Clustering is the technical method. Segmentation is the practical outcome.

For example, a clustering algorithm may group customers into five clusters. A marketing team may then interpret those clusters as:

• loyal customers,
• price-sensitive buyers,
• inactive users,
• high-value B2B customers,
• new low-engagement visitors.

The algorithm creates clusters. Humans turn them into meaningful segments.

Clustering and embeddings

Embeddings are numerical representations of content such as text, images, users, products or documents.

Clustering is often applied to embeddings because embeddings already represent similarity in vector space.

For example:

• text embeddings can be clustered into topics,
• image embeddings can be clustered into visual categories,
• product embeddings can be clustered into product families,
• user embeddings can be clustered into behaviour groups,
• prompt embeddings can be clustered into task types.

The quality of clustering depends on the quality of the embeddings. If the embedding model represents meaning poorly, the clusters will also be poor.

Clustering text data

Text clustering groups documents, queries, tickets, reviews or messages by similarity.

Common workflows include:

• convert text into vectors using TF-IDF or embeddings,
• choose a distance metric such as cosine similarity,
• apply a clustering algorithm,
• inspect representative documents in each cluster,
• name the clusters based on their content.

Text clustering is useful for topic discovery, support ticket analysis, keyword grouping, content audits and search evaluation.

For SEO, clustering can help group keywords by search intent or identify overlapping article topics.

Clustering image data

Image clustering groups images by visual similarity. This is usually done using image embeddings from a model rather than raw pixels.

For example, an e-commerce team may cluster product images to find similar products, duplicates, wrong category labels or unusual photos.

Image clustering can help with:

• catalogue organisation,
• duplicate detection,
• quality control,
• visual search,
• recommendation systems,
• dataset cleaning.

As with text, the embedding quality matters. Poor image representations create poor clusters.

Clustering customer data

Customer clustering is common in marketing, sales and analytics.

It can group customers by:

• purchase frequency,
• recency of last purchase,
• average order value,
• product preferences,
• discount sensitivity,
• email engagement,
• website behaviour,
• support history,
• company size or industry.

The result can support segmentation, personalisation, retention campaigns, lifecycle marketing and sales prioritisation.

But customer clusters must be interpreted carefully. A cluster is useful only if it leads to better decisions.

Clustering for anomaly detection

Clustering can help detect anomalies or outliers.

If most data points belong to clear clusters, points that do not fit any cluster may deserve investigation. These unusual points may indicate fraud, sensor failure, tracking errors, unusual customer behaviour or rare edge cases.

Some algorithms, such as DBSCAN, can explicitly mark points as noise. Other algorithms may require additional analysis, such as measuring distance from a cluster centre.

An outlier is not automatically bad. It may be an error, but it may also be the most important point in the dataset.

Clustering and dimensionality reduction

High-dimensional data can make clustering difficult. Distances may become less meaningful, and clusters may become harder to detect.

This is connected to the curse of dimensionality. When the number of dimensions is high, the feature space becomes sparse and local neighbourhoods become harder to trust.

Dimensionality reduction can help. Common methods include:

• PCA – a linear method often used before clustering,
• UMAP – useful for visualising high-dimensional structure,
• t-SNE – mainly used for visualising local neighbourhoods,
• autoencoder – a neural network that learns compressed representations.

However, clustering should not be judged only from a 2D projection. The original feature space still matters.

Clustering and feature selection

Feature selection can improve clustering by removing irrelevant or noisy variables.

If a dataset contains many weak features, the clustering algorithm may group points based on noise rather than meaningful similarity.

For example, customer clustering should not be dominated by random tracking parameters or temporary campaign IDs. Product clustering should not be dominated by internal database fields. Text clustering should not be dominated by boilerplate phrases.

Good feature selection can make clusters:

• more stable,
• easier to interpret,
• less noisy,
• more useful for business decisions,
• less sensitive to irrelevant variation.

Clustering and outliers

Outliers can strongly affect clustering.

In k-means, an outlier can pull a centroid away from the main group. In density-based methods, outliers may be identified as noise. In hierarchical clustering, outliers can create isolated branches.

Before clustering, it is useful to inspect:

• extreme values,
• data errors,
• duplicates,
• rare but valid cases,
• tracking artefacts,
• segment-specific behaviour.

Outliers should not always be removed. If the task is fraud detection or anomaly analysis, outliers may be the main target.

Clustering and data leakage

Data leakage can happen when clustering is used inside a machine learning pipeline.

For example, if clustering is applied to the whole dataset before train-test splitting, information from the test set may influence the cluster representation. This can make downstream evaluation too optimistic.

A safer process is:

• split the data first,
• fit preprocessing and clustering on the training data,
• apply the learned transformation to validation and test data where appropriate,
• evaluate downstream models on unseen data.

This is especially important when cluster labels become features for another model.

How to evaluate clustering

Clustering evaluation is difficult because there are no predefined correct labels in many cases.

Common evaluation approaches include:

• silhouette score – checks whether points are closer to their own cluster than to other clusters,
• Davies-Bouldin index – compares within-cluster similarity and between-cluster separation,
• Calinski-Harabasz index – compares between-cluster dispersion and within-cluster dispersion,
• stability testing – checks whether clusters remain similar across samples or settings,
• external labels – compares clusters with known labels if available,
• domain validation – checks whether clusters make practical sense.

No metric is perfect. A high score does not automatically mean the clusters are useful.

Cluster interpretation

After clustering, the next step is interpretation.

A cluster should be described using the original features and real-world context. For example, a customer cluster may be described by average order value, recency, product category, discount use and engagement level.

Useful questions include:

• What makes this cluster different?
• Which features are typical for this cluster?
• Is the cluster stable?
• Is it large enough to matter?
• Can it be named clearly?
• Does it support a useful action?
• Is it an artefact of preprocessing?

A cluster without interpretation is usually not useful for business decisions.

Clustering and visualisation

Clustering results are often visualised with scatter plots, PCA, UMAP or t-SNE.

Visualisation can help people understand groups, overlaps and outliers. It is especially useful when communicating results to non-technical stakeholders.

However, visualisation can mislead. A 2D projection may exaggerate separation, hide overlap or distort distances.

Good visualisation should:

• document the method used,
• show cluster labels clearly,
• avoid overinterpreting axes,
• include sample points or examples,
• compare with original features,
• avoid presenting exploratory plots as final proof.

Soft clustering vs hard clustering

Hard clustering assigns each point to exactly one cluster. K-means is a typical example.

Soft clustering assigns probabilities or degrees of membership. A point can belong partly to several clusters. Gaussian mixture models are a common example.

Hard clustering is simple and easier to communicate. Soft clustering is useful when boundaries between groups are not clear.

For example, a customer may be partly similar to a discount-driven segment and partly similar to a loyal high-value segment. Soft clustering can represent this uncertainty better than a strict assignment.

Exclusive, overlapping and hierarchical clusters

Different clustering approaches make different assumptions.

Exclusive clustering means each point belongs to one cluster.

Overlapping clustering allows a point to belong to more than one group.

Hierarchical clustering organizes clusters into levels, from broad groups to more detailed subgroups.

The right approach depends on the real-world structure. Some categories are naturally exclusive. Others overlap. Some need several levels of detail.

Clustering in recommendation systems

Clustering can support recommendation systems by grouping similar users, products or content.

For example:

• users can be clustered by behaviour,
• products can be clustered by similarity,
• articles can be clustered by topic,
• videos can be clustered by viewing patterns,
• queries can be clustered by intent.

These clusters can help generate recommendations, reduce search space, improve personalisation or explain why items are similar.

However, clustering alone is not a full recommendation system. It is usually one component.

Clustering in SEO and content strategy

Clustering can be useful in SEO because search queries and pages often form semantic groups.

For example, keyword clustering can group queries by intent:

• informational queries,
• comparison queries,
• transactional queries,
• problem-solving queries,
• brand queries,
• local queries.

Content clustering can also show whether a website has too many overlapping articles, missing topic coverage or weak internal linking between related pages.

For this use case, clustering should combine embeddings, search intent review and editorial judgment. Automatic clusters should not replace human SEO analysis.

Clustering in AI evaluation

Clustering can help evaluate AI systems by grouping prompts, documents, outputs or errors.

For example, a team may embed user prompts and cluster them to understand common task types. This can reveal whether the evaluation set covers real usage.

Clustering can also group model failures:

• factual errors,
• formatting errors,
• instruction-following errors,
• retrieval failures,
• unsafe outputs,
• tool-use mistakes,
• edge cases.

This is useful because model quality problems often appear in patterns, not as isolated failures.

Clustering and large language models

Large language models often work with embeddings and high-dimensional representations. Clustering can help inspect those representations.

For example, clustering can group:

• similar prompts,
• similar documents,
• similar retrieved passages,
• similar generated answers,
• similar failure cases,
• similar user intents.

This can support prompt engineering, evaluation design, retrieval tuning and content analysis.

Clustering and AI agents

In AI agent systems, clustering can help analyse user requests, tool calls, errors and workflows.

For example, logs from an agentic system can be clustered to find:

• common task types,
• repeated failure patterns,
• unusual tool-use behaviour,
• groups of prompts that need better handling,
• cases that should be routed to humans,
• possible misuse or prompt injection attempts.

This makes clustering useful not only for data science, but also for AI monitoring and governance.

Common mistakes with clustering

Clustering is easy to misuse because algorithms will produce groups even when the groups are not meaningful.

Common mistakes include:

• assuming clusters are real – an algorithm can create groups even in weak or noisy data,
• using irrelevant features – poor input creates poor clusters,
• not scaling data – large-scale variables may dominate distance,
• choosing k mechanically – the number of clusters needs interpretation, not only a metric,
• overinterpreting UMAP or t-SNE plots – visual separation is not proof,
• ignoring outliers – extreme values can distort clusters,
• forgetting business meaning – clusters must support decisions,
• using clustering as final truth – clustering is usually exploratory and should be validated.

When to use clustering

Clustering is useful when you want to discover groups and do not already have reliable labels.

It is a good choice when:

• you want to explore unknown structure,
• you need customer or product segments,
• you want to group similar documents or queries,
• you want to inspect embeddings,
• you want to find unusual groups or outliers,
• you need to simplify a complex dataset,
• you want to support downstream analysis or labelling.

Clustering is strongest as an exploratory method and as a way to organize complex data.

When not to use clustering

Clustering is not always the right method.

It may be a poor choice when:

• the data has no meaningful similarity structure,
• the features are noisy or irrelevant,
• the business goal requires known labels,
• the groups cannot be interpreted or used,
• the dataset is too small for stable grouping,
• the algorithm assumptions do not match the data,
• a simple rule-based segmentation would be clearer.

A clustering output is useful only if it helps answer a real question.

How to remember clustering

Clustering can be compared to sorting a box of mixed objects without being told the categories. You look for similarity and create groups: similar shapes, similar colours, similar sizes or similar functions.

A clustering algorithm does the same with data. It groups points that look similar under the chosen representation.

The important part is interpretation. The algorithm can form groups. People still need to decide whether those groups mean anything.

Related terms

• Machine learning – the broader field in which systems learn patterns from data and use them for prediction, classification or analysis.
• Unsupervised learning – machine learning where the model works with unlabelled data and tries to discover structure.
• K-means – a centroid-based clustering algorithm that groups data into a chosen number of clusters.
• DBSCAN – a density-based clustering method that can find irregularly shaped clusters and mark noise points.
• Hierarchical clustering – a clustering method that creates a tree-like structure of groups.
• Gaussian mixture model – a probabilistic clustering method that can assign soft cluster membership.
• Centroid – the centre point of a cluster, often used in k-means.
• Distance metric – a method for measuring how similar or different data points are.
• Embedding – a numerical representation of content or objects, often used for similarity search and clustering.
• UMAP – a nonlinear dimensionality reduction method often used to visualise high-dimensional data and clusters.
• t-SNE – a method used mainly for visualising high-dimensional data in two or three dimensions.
• Outlier – a point that appears unusual compared with the rest of the data.
• Curse of dimensionality – a problem where high-dimensional spaces become sparse and harder to learn from reliably.
• Feature selection – selecting useful original variables and removing irrelevant or redundant ones.
• Silhouette score – a metric used to evaluate how well points fit within their assigned clusters.
• Customer segmentation – grouping customers into meaningful segments for analysis, marketing or sales.

Sources and further reading

• Clustering – scikit-learn.org – June 2026 – explains clustering methods available in scikit-learn, including k-means, DBSCAN, hierarchical clustering, spectral clustering and other algorithms.
• Comparing different clustering algorithms on toy datasets – scikit-learn.org – June 2026 – visually compares how different clustering algorithms behave on different data shapes.
• Demo of DBSCAN clustering algorithm – scikit-learn.org – June 2026 – shows how DBSCAN finds dense regions and expands clusters from core samples.
• What is k-means clustering? – developers.google.com – June 2026 – explains k-means as a clustering algorithm that groups data points by minimizing distance to cluster centroids.
• Clustering algorithms – developers.google.com – June 2026 – describes centroid-based, density-based and distribution-based clustering approaches.
• Advantages and disadvantages of k-means – developers.google.com – June 2026 – explains strengths and limitations of k-means, including scalability, sensitivity and high-dimensional challenges.
• What is clustering? – ibm.com – June 2026 – defines clustering as an unsupervised learning method that groups objects or observations based on similarities or patterns.
• What Is Unsupervised Learning? – ibm.com – June 2026 – explains unsupervised learning and its use for discovering hidden groupings in unlabelled data.
• What is k-means clustering? – ibm.com – June 2026 – explains k-means clustering as an unsupervised algorithm for grouping unlabelled data points.
• silhouette_score – scikit-learn.org – June 2026 – documents a common metric for evaluating clustering quality based on within-cluster and between-cluster distances.