A Hybrid Denoised Framework for Customer Intelligence: Sequencing Density-Based Noise Filtering and AHP-Weighted RFM Segmentation

Retailers increasingly rely on customer segmentation to guide marketing spend, allocate service resources, and anticipate churn, yet many segmentation pipelines feed raw or lightly cleaned transaction data straight into a clustering algorithm, which tends to blur cluster boundaries and undermine interpretability. This paper introduces a framework in which Density-Based Spatial Clustering of Applications with Noise (DBSCAN) is deliberately repurposed away from its usual role as the terminal clustering algorithm and instead used as an upstream filter that screens out multivariate noise before clustering ever begins — a sequencing that, to our knowledge, has not been systematically tested inside an AHP-weighted RFM-K-Means pipeline. The pipeline moves through four stages: Recency-Frequency-Monetary (RFM) feature construction on the UCI Online Retail dataset (397,884 transactions, 4,338 customers); log-transform and Min-Max rescaling of the three RFM dimensions; differential weighting of those dimensions using Analytic Hierarchy Process (AHP) judgments (wR = 0.15, wF = 0.35, wM = 0.50, CR = 0.021); and, finally, K-Means clustering on the resulting denoised, weighted feature space. Rather than reporting only the combined pipeline's performance, a controlled three-way ablation (Models A, B, and C) isolates the marginal contribution of AHP weighting from the additional marginal contribution of DBSCAN denoising, so the central empirical claim — that the ordering itself, filter for noise, then weight, then cluster, produces measurable gains over both a plain K-Means baseline and an AHP-weighted variant that skips denoising — rests on evidence for each step individually rather than on the pipeline as a whole. The full pipeline (referred to below as Model C) reaches a Silhouette Score of 0.3872, a Davies-Bouldin Index of 0.9181, and a Calinski-Harabasz score of 4,498.10, having flagged and removed 10.8% of customer records as multivariate noise beforehand. Clustering the retained customers yields three segments — High-Value (847 customers, 49.1% of revenue), Regular (1,706 customers, 37.3%), and At-Risk (1,317 customers, 13.5%) — and bootstrap resampling together with a permutation test confirm that these gains are not artifacts of a lucky initialization. Taken together, the results show that density-based filtering, treated as a distinct preprocessing step rather than a substitute for K-Means, provides a repeatable and statistically validated way to improve both cluster quality and the business usefulness of the resulting segments.