Customer segmentation that changes decisions
How clustering can reveal useful customer or store groups without turning the analysis into decorative personas.
Clustering is an exploratory tool
K-means, hierarchical clustering, mixtures, and density-based methods organize similar observations. They do not discover objective customer species. A good segment is stable enough to understand and different enough to change an action.
Representation decides the result
Recency, frequency, margin, category mix, channel behavior, geography, and lifecycle often matter more than the clustering algorithm. Scaling, transformations, missingness, and time windows can completely change the groups.
Validate with behavior
Useful segments predict a difference outside the variables used to build them: response to a promotion, churn risk, support cost, product affinity, or service capacity. If nobody can act differently, the segmentation is unfinished.