Reducing Model Complexity in Bank Customer Churn Prediction Using Dimensionality Reduction and Explainable Machine Learning

Abstract

Customer churn poses significant challenges to banking profitability, with acquiring new customers costing 5-7 times more than retaining existing ones. Traditional statistical methods struggle with high-dimensional data and complex non-linear patterns. This study addresses these limitations by integrating advanced dimensionality reduction techniques with optimized machine learning to achieve accurate, interpretable churn prediction. Three dimensionality reduction techniques—UMAP, NCA, and PLSDA—were evaluated using a 10,000-customer European banking dataset across ten machine learning classifiers: SVM, Logistic Regression, KNN, Gaussian Naive Bayes, Decision Tree, Random Forest, AdaBoost, Bagging, Stacking, and Voting. Data preprocessing included SMOTE balancing, StandardScaler normalization, and one-hot encoding. GridSearchCV optimized hyper parameters systematically. Dual evaluation employed 80-20 train-test split and 10-fold cross-validation. SHAP framework provided comprehensive explainability through five visualization techniques. PLSDA emerged as the superior approach, achieving 84.68% accuracy with Random Forest using 8 features—20% dimensionality reduction while retaining 99% performance. Cross-validation confirmed robustness: 85.27 ± 0.50% accuracy, 92.79 ± 0.70% ROC-AUC. Ensemble methods outperformed single classifiers consistently. SHAP analysis identified Age, Number of Products, and Balance as dominant churn drivers. This study demonstrates that PLSDA-optimized machine learning achieves competitive accuracy with reduced computational complexity and enhanced interpretability. The framework provides actionable insights for targeted retention strategies while meeting regulatory compliance requirements for practical banking implementations.