Investigating clustering techniques to establish target client groups efficiently

This study develops a decision-oriented framework for comparing customer-segmentation algorithms rather than treating technical benchmarking as a contribution in itself. K-Means, Agglomerative Hierarchical Clustering, DBSCAN, and Affinity Propagation were applied to the public UCI Online Retail dataset from 1 December 2010 to 9 December 2011. The raw file contains 541,909 transactions; after predefined cleaning, 406,829 valid transactions were aggregated into 4372 customer profiles represented by Recency, Frequency, Monetary value, and Product Variety. The comparison combined internal validity, resampling stability, matched runtime, segment-size balance, profile distinctness, and deployment feasibility. Deployment feasibility was treated as decision support, not as an observed business outcome. Affinity Propagation achieved the strongest internal structure (Silhouette = 0.65±0.01; Davies-Bouldin = 0.60±0.02), but produced 45 fragmented clusters, required the longest runtime (14.60±0.83 s), and showed moderate stability (ARI = 0.69). K-Means did not have the best internal-validity score, but achieved the highest stability (ARI = 0.82), the shortest runtime (1.84±0.12 s), ten comparatively balanced clusters, and direct centroid-based reassignment. It is therefore preferred only when repeatable and manageable operational segmentation is required; Affinity Propagation is preferable when structural separation is the main objective. Hierarchical clustering supports exploratory sub-segmentation, whereas DBSCAN supports density and noise diagnosis. The contribution is an explicit selection protocol separating structural validity, deployment feasibility, and business validation. Because the dataset is historical and contains no campaign, retention, customer-lifetime-value, or revenue-uplift outcome, all CRM actions remain hypotheses requiring temporal holdout tests or field experiments.