Explainable AI Dimensionality Reduction Techniques Based for Client Attrition Prediction

Abstract

In the financial field, predicting client churn is of great significance for banks to maintain profitability and reputation, and attract new customers. Traditional prediction methods, such as KNN and RF, typically assume linear relationships and struggle with complex linear patterns in customer behavior. Recent research has employed machine learning and dimensionality reduction techniques to improve prediction accuracy, but most studies have focused on a single model and lack a comprehensive analysis of how different dimensionality reduction methods interact with various classifiers. This study employed three dimensionality reduction techniques: Uniform Manifold Approximation and Projection (UMAP), Neighbourhood Component Analysis (NCA), and Partial Least Squares Discriminant Analysis (PLSDA) to combine 10 machine learning classifiers including SVM, KNN, logistic regression, etc., to conduct research on the prediction of credit card customer churn. The Bank Churners dataset was balanced by using SMOTE method. Additionally, each classifier is optimized using GridSearchCV and its performance is evaluated using key evaluation metrics. The combined use of the reduced dataset from NCA with the K-nearest neighbor classifier achieved the best performance, with an accuracy rate of 96.21% and a precision rate of 95.84%. Among all the dimensionality reduction methods, NCA achieved excellent results in most of the classifiers. This study systematically compared the three-dimensional reduction strategy and ten classifiers for credit card churn prediction. The research results emphasized that NCA and KNN formed an effective combination for building an accurate and interpretable churn prediction system. This study provides practical insights for financial institutions aiming to deploy efficient customer retention models.