Dynamic RFM Model Selection for Customer Segmentation Using Meta-Learning

Abstract

The Recency-Frequency-Monetary (RFM) model has been extended through numerous formulations to accommodate the varying characteristics of customer transaction datasets. Weighted, entropy-based, logarithmic, fuzzy, and hybrid RFM formulations have been proposed for different data conditions. Selecting the most appropriate formulation for a given dataset, however, continues to depend on empirical evaluation or domain expertise. In this paper, we propose a Dynamic RFM Model Selection (DRMS) framework that recommends the most appropriate RFM formulation for a given transaction dataset. The framework explicitly separates dataset characterisation from RFM model selection. A compact set of seven behavioural and business descriptors characterises the dataset, which a meta-learning engine uses to recommend the most suitable formulation. Experimental results on ten benchmark datasets confirm that no single RFM formulation performs consistently well across all transaction datasets. The proposed Random Forest meta-learner achieves 90% recommendation accuracy under leave-one-out cross-validation, outperforming all three baseline meta-learning models. An ablation study confirms that Monetary Skewness, Attribute Balance, and Information Imbalance are the most informative descriptors for formulation recommendation.