Hard vs. Soft Clustering for RFM Segmentation: A Joint Evaluation of Method Suitability Across Singular Transaction and Subscription Business Models

Hard vs. Soft Clustering for RFM Segmentation: A Joint Evaluation of Method Suitability Across Singular Transaction and Subscription Business Models

Warwick Business School

Ying Ting Chan

2025

Purpose – This study examines the suitability of hard versus soft clustering approaches for customer segmentation across distinct business model types. Specifically, it evaluates whether homogeneous marketing treatment implied by K-means is justified in subscription and singular transaction contexts, and whether Fuzzy C-means (FCM) offers a more appropriate representation of behavioural overlap. Methodology – Using the RFM (Recency–Frequency–Monetary) framework, the study applies both K-means and FCM to two datasets: the Online Retail dataset (transaction model) and the Customer Churn dataset (subscription model). RFM metrics were adapted for subscription contexts to reflect product interactions rather than discrete purchases. Clustering performance was assessed through the silhouette score (K-means) and fuzzy partition coefficient (FCM), following a consistent CRISP-DM process for data preparation, modelling, and evaluation. Findings – Results indicate a reversal of theoretical expectations. The subscription dataset exhibited considerable behavioural overlap, with weak cluster separation, challenging the assumption of homogeneous treatment. Conversely, the transaction dataset displayed moderately strong partitions, suggesting that standardisation of product attributes may reduce behavioural heterogeneity. These findings highlight that clustering suitability depends not only on business model type but also on contextual factors such as product standardisation and tier design. Originality – This study is the first to align the selection of hard versus soft clustering methods with business model typologies and to adapt RFM metrics for subscription contexts. By integrating FCM as a diagnostic tool for testing the validity of homogeneous treatment assumptions, it addresses a critical gap in behavioural segmentation research and provides a practical decision framework for guiding clustering choice.

Customer segmentation, Machine learning, Clustering, Business analytics, Customer behaviour, Business model