Clustering searches for recurring groups in data without assigning the groups in advance. An algorithm will always divide the records somehow, even when the resulting clusters have no stable business meaning. Marketways tests whether the groups persist, can be understood and require different actions. The client gains a useful basis for differentiated offers, service or investigation only when those tests are met.
The decision this method supports
We use Clustering & Unsupervised Learning to help clients answer: Does the data contain stable and useful structure that was not defined in advance?
How the method works
Unsupervised learning searches for structure without receiving a correct label for each observation. Clustering groups similar records, while representation methods reduce complex information to a smaller set of dimensions. The discovered structure is a hypothesis until business meaning and stability are established.
A business example
A retailer may discover several purchasing patterns in transaction data. The clusters become useful only if the groups remain reasonably stable and lead to different decisions about offers, service or inventory.
How the client uses the result
Clustering can reveal recurring groups that were not defined in advance. When the groups are stable and actionable, a business can design different offers, service levels or investigations for cases that should not be treated alike.
What we deliver
We produce predictions, scores, groups, alerts or extracted information together with validation evidence. A manager needs operating thresholds, error consequences, escalation rules and a plan for monitoring change after deployment.
Limits and complementary methods
An algorithm will produce groups even when no useful business segments exist. Stability, interpretation and a different action for each group must be demonstrated.
Selected methods and techniques
We select from these established methods according to the decision, evidence and operating conditions.
- Clustering: Find groups of similar cases without deciding the categories beforehand. Examine whether the groups are stable and meaningful before using them for decisions.
- Cohort analysis: Compare groups that share a defined starting event, period or characteristic and follow the same outcome over comparable relationship ages. State the cohort rule, time origin, eligible population, exposure and observation cutoff; align results by tenure when that is the question; show the number still observed and at risk at each time; and separate tenure effects from calendar events, changing acquisition mix and incomplete follow-up. Cohort analysis describes how group histories differ and can reveal when retention changes. It does not by itself identify the cause of the difference, predict an individual's departure or estimate the effect of a retention action.
- Exploratory factor analysis: Look for underlying dimensions that could explain why survey answers vary together, without fixing their structure in advance. The dimensions need interpretation and checking. Confirmatory factor analysis instead tests a structure specified before examining the results.
- Issue clustering: Group issue records by similarity without starting from a complete set of predefined outcome labels. Define the record and corpus, period, text or behavioural features, preprocessing, similarity measure or model, treatment of duplicates and rare cases, and the rules for choosing the number of clusters and assessing their stability. Inspect representative and borderline records with subject-matter reviewers, test stability across samples or time and retain an unclassified route where patterns are weak. Clustering discovers candidate groupings; it does not by itself create a governed taxonomy, assign a verified cause or estimate issue prevalence outside the collected records.
- Latent-class / segment preference analysis: Find customer groups from observed preference patterns when no labels are supplied. Estimate those groups from their choices, allowing uncertainty about which group a person belongs to.
- Latent-class analysis: Estimate unobserved groups from recurring patterns in categorical, choice or response data using a probabilistic model. Define the variables and population, compare plausible numbers and forms of classes, examine fit and stability, and report each case's membership probabilities rather than forcing certainty where classes overlap. Describe and validate the classes on evidence not used merely to name them where possible. The classes depend on the model, measures and sample; they are not automatically natural or causal groups. Their commercial use also requires meaningful size, reproducible assignment and different needs or responses.
- PCA: Combine many related numerical measures into fewer summary measures called principal components, retaining as much of their variation as possible. PCA (principal component analysis) is a way to simplify the data, not proof that the summaries represent important business concepts or predict the outcome. The scales of the original measures affect the result.
- RFM analysis: Calculate recency, frequency and monetary measures for a defined customer population using explicit transaction rules, observation windows, reference date and treatment of returns or missing activity. Convert the measures into stated scores or segment rules and test whether results remain useful when cut-offs or periods change. RFM summarises historical purchasing behaviour; it does not estimate future CLV, explain why behaviour occurred or show that an intervention will work.
- Segmentation: Divide a defined population into groups that are useful for understanding or action, using explicit variables, a stated method and reproducible assignment rules. For customer-experience work, groups may differ in needs, journey, behaviour, satisfaction or response to service. Document data preparation, model or rule selection, segment size, stability, uncertainty and validation; describe how new cases are assigned; and check that sensitive attributes and very small groups are handled responsibly. Segments simplify variation rather than creating natural truths. They are distinct from informal personas, and experience segmentation does not by itself estimate market demand, attractiveness or causal treatment response.
- Segmentation / clustering: Divide a defined population into groups for a stated analytical or decision purpose. Segmentation can apply rules chosen in advance, while clustering estimates groups from similarity patterns without predefined labels. State the population, variables, scaling, grouping rule or model, and test stability, separation, interpretability, size and usefulness on data beyond the construction sample. The groups depend on these choices; they are not automatically natural, causal, permanent or suitable for decisions beyond the evidence used to create them.
Parent method family
Machine Learning & Predictive Analytics explains how this method connects to adjacent methods and relevant services.
Related service families
These service families contain business questions supported by this method. Service pages link to the wider method family so readers can understand the complete analytical approach.