Adaptive multidimensional rebalancing for customer churn prediction in marketing analytics

Background Customer churn prediction is a strategic imperative for maximizing Customer Lifetime Value (CLV) and optimizing retention return on investment (ROI). However, real-world churn datasets suffer from severe class imbalance and high-dimensional class overlap, causing conventional models to ignore minority patterns. Existing resampling techniques often generate synthetic noise in ambiguous deci-sion boundaries, degrading both predictive accuracy and business value. Methods This study proposes the Adaptive Multidimensional Rebalancing (AMR) framework that dynamically evaluates the local topological structure of the feature space, specifically local density, class overlap, and feature variability, to adap-tively allocate synthetic minority samples. AMR employs a constrained interpo-lation mechanism that prevents synthetic instances from bleeding into majority-class regions. The framework is evaluated across four diverse business sectors (Telecommunications, Banking, Credit Card, and Insurance) and integrated with cost-sensitive ensemble classifiers. Results Extensive experiments demonstrate that AMR achieves statistically sig-nificant improvements in predictive stability (AUPRC) compared to state-of-the-art baselines like SMOTE-ENN ( p < 0.05). AMR exhibits superior computational ef-ficiency, reducing resampling runtime by approximately 35% compared to hybrid methods. By maximizing the Expected Maximum Profit (EMP) and leveraging SHAP-based Explainable AI (XAI), AMR enhances predictive sensitivity and pro-vides actionable, interpretable insights. Conclusions AMR bridges the gap between complex machine learning and profit-centric marketing objectives, offering a robust, scalable, and interpretable solution for customer retention.