Supervised Distance Metric Learning with RFM-Based Business Knowledge for Customer Segmentation

The effectiveness of customer segmentation in e-commerce depends on the ability to identify meaningful patterns in customer behavior and group similar customers together. This study explores supervised distance metric learning for customer segmentation using Large Margin Nearest Neighbor (LMNN) with Recency-Frequency-Monetary (RFM) labeling based on business expertise. Using the UCI Online Retail dataset comprising 4,338 customers characterized by 11 extended RFM features, we compare metric learning approaches to PCA-based and normalized baselines. Experimental results demonstrate that LMNN+K-Means achieves superior clustering quality (Silhouette Score: 0.363, Calinski-Harabasz: 1,394) and competitive Classification accuracy (97.56% on a single seed, 96.135% on average over 10 independent runs). These findings indicate that RFM labeling based on business expertise helps reduce the circularity problem inherent in pseudo-label-based approaches, while producing actionable customer segments aligned with business knowledge.