Tackling Customer Segmentation
Our client came to us asking how to group seventeen thousand customers so it could set prices by group. The obvious grouping — the industry label each customer already carries — turns out to predict almost nothing about how a customer actually behaves. This is the method that finds the groups that do: clustering on mixed numeric and categorical data with KPrototypes, choosing the number of groups with the elbow method, and the finding that the legacy categories cut across every discovered tier.