Cluster analysis – a guide to smarter data-driven decisions

Adobe for Business Team

02-24-2025

A smiling man in a patterned shirt talks on the phone while holding a tablet in a retail store. Floating elements show a personalised ad segment for 11K seasonal shoppers and social media posts featuring green fashion items.

Cluster analysis is a statistical method for grouping similar data points together and distinguishing them from other groups.

Picture a clothing retailer grouping customers by purchasing habits – frequent buyers, seasonal shoppers, or one-time purchasers. Cluster analysis helps businesses identify these groups and tailor marketing strategies, from targeted ads to personalised offers.

In marketing, cluster analysis segments consumers into distinct groups based on shared characteristics. This helps businesses better understand their target audience and tailor their marketing strategies accordingly.

What you’ll learn:

What is cluster analysis, and how does it work?

Cluster analysis is a form of unsupervised classification – meaning there are no predefined classes, definitions, or expectations going in. It’s a statistical data mining technique that groups observations which are similar to each other but distinct from other groups.

A helpful way to understand clustering is to picture someone sorting chocolates from a sampler box. They may have preferences for certain types of chocolate.

As they sort through the box, there are plenty of ways to group the chocolates – by milk versus dark chocolate, with or without nuts, with or without nougat, and so on.

Sorting chocolates into groups based on shared characteristics – that’s clustering. It’s something we all do naturally.

For instance, an ecommerce platform might group customers by purchasing habits – such as budget-conscious shoppers, premium product buyers, and occasional browsers. This segmentation lets the platform create tailored promotions for each group, driving engagement and sales.

Understanding cluster analysis

Cluster analysis sits at the heart of data analysis. It’s easy to see why fields like finance, insurance, retail, ecommerce, and marketing rely on it to identify patterns and relationships within their data.

There are five main clustering approaches, with k-means clustering and hierarchical (or hierarchy) clustering being the most common. The approach an organisation takes depends on what’s being analysed and why. Using visualisation techniques such as scatter plots and dendrograms, businesses can present their cluster analysis results clearly and in a way that’s easy to understand.

What is the purpose of clustering datasets?

The core purpose of cluster analysis in marketing is to build groups or clusters where the observations within each group are as similar as possible.

Ultimately, the goal depends on the application. In marketing, clustering helps marketers uncover distinct customer groups within their base – and put that knowledge to work building targeted campaigns.

For instance, clustering might help an insurance company identify groups of motor insurance policyholders with a high average claim cost.

How a company uses clustering depends on its industry, business unit, and the objectives it’s looking to achieve.

Why does cluster analysis matter for business strategy

Cluster analysis can benefit a company in a number of ways, including how they market their products.

It shapes who they market those products to, which retention and sales strategies to employ, and how to assess prospective customers.

By clustering current customers, companies can assess their lifetime value against their likelihood of churning. This informs how they engage different customer segments and how to identify new high-value prospects.

What are the different types of clustering and when do you use them

A row of five green icons representing different types of clustering algorithms: partitioning, hierarchical, density-based, grid-based, and model-based algorithms.

There are five major clustering algorithms:

Partitioning algorithms
Hierarchical algorithms
Density-based algorithms
Grid-based algorithms
Model-based algorithms

Clustering algorithm

Description

Best for

Disadvantages

Marketing use case

Partitioning algorithms

Partitioning algorithms, such as k-means clustering, divide the dataset into a predefined number of clusters by optimising an objective function (e.g., minimising the sum of squared distances).

Best suited to datasets where the number of clusters is known upfront and the clusters are clearly separated.

Requires specifying the number of clusters upfront
Can struggle with clusters of varying sizes and densities
Sensitive to outliers

Segmenting corporate clients into distinct groups based on purchase patterns, enabling targeted B2B email campaigns and personalised product offerings.

Hierarchical algorithms

Hierarchical algorithms, including agglomerative and divisive clustering, build a nested hierarchy of clusters by merging or splitting them based on similarity.

Useful when the underlying data has a hierarchical structure or when the number of clusters isn’t known in advance.

Computationally intensive for large datasets
Early clustering decisions cannot be undone
Sensitive to noise and outliers

Organising business customer data into a hierarchical structure (e.g., by industry, then by company size) to tailor multi-level marketing strategies and account management.

Density-based algorithms

Density-based algorithms, like DBSCAN, identify clusters as dense regions of data points separated by lower-density areas, enabling the discovery of clusters of any shape.

Effective for datasets with clusters of varying shapes and sizes, particularly when noise is present.

Sensitive to parameter selection
Can struggle with clusters of varying densities and high-dimensional data
Can misclassify border points

Detecting clusters of high engagement among B2B clients within noisy transaction data to focus marketing efforts on high-value accounts or areas of concentrated business activity.

Grid-based algorithms

Grid-based algorithms divide the data space into a finite number of cells forming a grid structure, then identify clusters based on the density of data points within those cells.

Efficient for large datasets where a fast clustering method is required.

Heavily dependent on grid resolution
May not capture clusters of arbitrary shapes
Potential loss of detail

Quickly clustering large volumes of B2B lead data (e.g., by geolocation or other business attributes) to pinpoint regional hotspots for targeted sales outreach and marketing campaigns.

Model-based algorithms

Model-based algorithms assume that the data is generated by a mixture of underlying probability distributions, with the aim of estimating the parameters of those distributions.

Best suited to datasets where the data distribution can be accurately modelled by statistical distributions.

Computationally expensive
Requires correct model assumptions
Sensitive to initial conditions and potential model misspecification.

Segmenting B2B customers by fitting models (such as Gaussian mixture models) to transaction data, uncovering distinct buying behaviour segments for tailored marketing strategies.

style

grid width 12

What are the characteristics of a good cluster analysis?

A good cluster analysis accurately groups data in a way that’s useful and actionable. It uncovers real patterns in the data, leading to insights that drive decisions. A bad cluster analysis, on the other hand, creates misleading or arbitrary groups that don’t solve a problem or add value.

The characteristics of good cluster analysis are:

Real-world usefulness: The clusters align with meaningful differences in behaviour, preferences, or traits that can be acted upon.
Clear group separation: Each cluster is distinct, making it easy to interpret and apply.
Consistency: The same input data should produce similar clusters every time.
Scalability: The cluster analysis works effectively even as the dataset grows.
Actionable insights: The clusters provide a clear direction for marketing, product development, or operations.

For example, imagine you’re segmenting B2B customers based on their buying habits. A good clustering model groups them into:

High-volume buyers: Who need bulk discounts and priority service
Seasonal buyers: Who respond well to time-sensitive promotions
Low-frequency buyers: Who may need more nurturing and education

Each group can be targeted with a specific marketing strategy, improving conversions and increasing customer satisfaction.

In contrast, poor cluster analyses offer:

Arbitrary clusters: The groups don’t represent meaningful patterns.
Poor separation: Different clusters overlap too much to be useful.
Overfitting: The model forces patterns that don’t exist.
Inconsistent results: Running the same analysis multiple times leads to different clusters.
No business application: The clusters don’t help answer any useful question.

Consider a marketing team that clusters customers based on the number of vowels in their company name. While mathematically possible, this grouping has zero business value – it doesn’t predict behaviour, preferences, or needs. The result? A useless segmentation that wastes time and resources.

What are the disadvantages of cluster analysis, and how can companies avoid problems?

Disadvantage

Problem

How to avoid it

Arbitrary number of clusters

Determining the optimal number of clusters (k) can be challenging and may not reflect the true data structure.

Use methods like the elbow method or silhouette score to estimate the appropriate k. Experiment with different values and validate the results.

Sensitivity to outliers and noise

Outliers can distort cluster formation, leading to inaccurate groupings.

Pre-process your data to identify and handle outliers. Consider density-based clustering algorithms such as DBSCAN, which are more robust to noise.

Poor interpretability

Clusters may be difficult to understand or apply in practical scenarios.

Carefully select relevant features. Use visualisation techniques like Principal Component Analysis (PCA) or t-Distributed Stochastic Neighbour Embedding (t-SNE) to improve interpretability.

Unequal cluster sizes and density

Algorithms like k-means assume clusters of similar size and density, which may not be realistic.

Use alternative methods such as hierarchical clustering or Gaussian mixture models (GMM) that can accommodate clusters of varying shapes and sizes.

Computational complexity

Clustering large datasets can be resource-intensive and time-consuming.

Implement grid-based or sampling-based approaches to improve computational efficiency.

Overfitting to noise

The model might identify patterns in random noise, resulting in spurious clusters.

Regularly validate clusters against real-world business logic and use holdout datasets to check for overfitting.

Dependency on feature selection

Poor feature selection can result in misleading clusters.

Apply thorough feature selection or use dimensionality reduction techniques such as PCA or linear discriminant analysis (LDA) before clustering.

style

grid width 8

Companies can avoid clustering pitfalls by:

Validating results: Compare clusters with real-world data and confirm they make practical sense.
Testing different algorithms: Try various clustering methods to find the most suitable one for your data.
Carrying out regular updates: Update your clustering analysis periodically as data changes to keep results accurate.
Using domain expertise: Work with subject matter experts to ensure clusters are meaningful and actionable.

How do you perform cluster analysis?

A step-by-step flowchart with six green circles showing the clustering analysis process: choosing an analysis method, determining the number of cases, selecting variables, deciding on standardisation, applying an algorithm, and finalising clusters.

Step 1: Choose an analysis method

The first step in cluster analysis is choosing the analysis method, which depends on the size of the data and the types of variables.

Hierarchical clustering suits small datasets, for example, while k-means clustering works better for moderately large datasets where the number of clusters is known in advance.

Large datasets typically involve a mix of variable types and generally require a two-step procedure.

Step 2: Determine the number of cases

Once you’ve decided on an analysis method, begin by choosing the number of cases to divide into homogeneous groups or clusters. Those cases, or observations, can be any subject, person, or thing you want to analyse.

Step 3: Select variables for analysis

Next, choose which variables to include. There could be 1,000 variables, or even 10,000 or 25,000. The number and types of variables selected will determine which algorithm to use.

Step 4: Decide on variable standardisation

Then decide whether to standardise those variables so that each one contributes equally to the distance or similarity between cases. The analysis can be run with both standardised and unstandardised variables.

Step 5: Apply the chosen algorithm

Each analysis method takes a different approach.

For k-means clustering, select the number of clusters. The algorithm then iteratively estimates cluster means and assigns each case to the cluster with the smallest distance to its mean.
For hierarchical clustering, choose a statistic that measures how far apart or how similar two cases are.

Step 6: Finalise the number of clusters

Finally, the algorithm selects a method for forming groups and determines how many clusters are needed to represent the data. It assesses how similar the clusters are and splits them accordingly.

What do you do with the results of a cluster analysis?

Depending on the clustering method, there’s usually an associated visualisation – a common way to investigate results. For k-means clustering, x and y axes are typically used to show the distance between groups of observations.

This type of visualisation makes the groupings clearly apparent. For hierarchical clustering, a visualisation called a dendrogram is used, which shows the splits in the cut tree.

How do you make sure your cluster analysis is accurate?

First, assess cluster tendency. Before applying any clustering algorithm, check whether your dataset has the potential to form meaningful clusters or is simply randomly distributed.

One common approach is the Hopkins statistic, which measures how likely it is that your data is uniformly distributed. A value near 0 suggests a strong cluster tendency, while a value around 0.5 indicates randomness.

Visual tools like the visual assessment of cluster tendency (VAT) also help, by reordering the dissimilarity matrix to highlight potential clusters visually. If these tests confirm that your data naturally groups together, you can proceed with clustering. If not, clustering may not yield useful insights.

Next, determine the optimal number of clusters. Selecting the right number of clusters (k) is critical – too few may oversimplify the data, while too many can lead to overfitting.

The elbow method is one popular approach: Plot the within-cluster sum of squares against the number of clusters and look for the point where clustering performance improvements begin to level off – the ‘elbow.’ Another useful metric is the silhouette score, which evaluates how well each data point fits into its assigned cluster relative to the others. Higher silhouette scores indicate more distinct, well-separated clusters.

The gap statistic also helps by comparing observed within-cluster dispersion against what’s expected under a null distribution, making it easier to identify the optimal k where the gap is largest.

Finally, evaluate clustering quality. Once clusters are established, confirm that they are both internally cohesive and externally well-separated.

The silhouette coefficient, which ranges from −1 to 1, is widely used for this purpose – a score closer to 1 means the clusters are well defined. The Dunn index calculates the ratio between the smallest inter-cluster distance (between observations not in the same cluster) and the largest intra-cluster distance (within a cluster). Higher Dunn index values indicate better quality clusters. By contrast, the Davies–Bouldin index measures the average similarity between each cluster and its most similar one, with lower values indicating better clustering quality.

In summary, accurate cluster analysis comes down to a three-step process:

Assess cluster tendency: Determine whether your data naturally forms clusters using statistical tests like the Hopkins statistic and visualisation tools such as VAT.
Determine the optimal number of clusters: Identify the right number of clusters (k) using methods such as the elbow method, silhouette score, and gap statistic to avoid oversimplification or overfitting.
Evaluate clustering quality: Confirm that your clusters are both compact and well separated using metrics such as the silhouette coefficient, Dunn index, and Davies-Bouldin index.

Getting started with cluster analysis

The core benefit of cluster analysis is that it helps businesses uncover patterns and relationships within their data, enabling them to make informed decisions and act on real-time insights.

Ready to get started with cluster analysis? The first step is finding a proven software tool that can help you analyse and interpret your data effectively.

Adobe Analytics turns real-time data into real-time insights. More than a web analytics solution, it collects data from every point in the customer journey and turns it into insights that guide your next best action. Analytics uses artificial intelligence (AI) to deliver predictive insights based on the full scope of your data, giving users the ability to view and manipulate data in real time.

Request a demo or watch the overview video to learn more about Adobe Analytics.