An Explainable Machine Learning Approach With Density-Regularized Oversampling for Industrial Customer Churn Prediction

Abstract

Customer churn prediction has become a vital statistical endeavour for businesses seeking to keep clients and increase their profitability over time. Although numerous data-driven models have been proposed, churn prediction remains difficult due to complex customer behaviours, heterogeneous attributes, and high overlap between churners and non-churners. Existing studies proposed various machine learning (ML) and deep learning (DL) models along with improved feature engineering. However, the fundamental research gap is a lack of attention to class imbalance handling techniques that retain inherent data distribution in churn prediction tasks. More specifically, conventional oversampling approaches frequently generate synthetic instances without taking into consideration local data density, resulting in noisy samples and poor classifier performance. This research suggests a density-regularized SMOTE-ENC framework that specifically incorporates neighbourhood-based density restrictions into the process of creating synthetic samples for mixed-type churn datasets. The proposed method selectively retains synthetic minority samples that lie closer to minority class manifolds while being sufficiently distant from majority class regions. The proposed density-regularized oversampling framework consistently improves churn prediction performance across all datasets, achieving 84.00% accuracy, 60.16% precision, 64.46% recall, and 62.14% F1-score on Bank Churn; 77.44% accuracy, 55.38% precision, 77.21% recall, and 64.49% F1-score on Telco-IBM; and 95.65% accuracy, 87.46% precision, 87.24% recall, and 87.09% F1-score on Retail Churn. Compared with the original imbalanced datasets without oversampling, the proposed method improves the F1-score by 7.3%, 3.1%, and 0.5% on the Bank, Telco-IBM, and Retail datasets, respectively.