Investigating the Impact of Machine Learning Algorithms on Enhancing the Quality of Customer Experience at Bank Mellat

Abstract

This study aimed to investigate the impact of machine learning algorithms on enhancing the quality of customer experience at Bank Mellat by evaluating their predictive effectiveness, robustness, interpretability, and operational applicability in major banking processes. This applied, quantitative, descriptive-explanatory, and predictive study used anonymized operational and transactional data from Bank Mellat during the second quarter of 2023. After data cleaning, removal of incomplete and outlier observations, normalization, aggregation, and feature engineering, 8,500 valid records were retained and divided into training, validation, and test sets using a 70:15:15 ratio. Ten-fold cross-validation was employed to reduce overfitting. Logistic regression, decision tree, random forest, support vector machine, XGBoost, multilayer perceptron, Isolation Forest, LSTM, Transformer, and hybrid LSTM-XGBoost models were evaluated for credit-risk prediction, fraud detection, and foreign-exchange service-demand forecasting. Model robustness was assessed through sensitivity and crisis-scenario analyses, while SHAP, LIME, and counterfactual explanations were used to evaluate interpretability. Statistical comparisons among algorithms were conducted using the Friedman and Wilcoxon signed-rank tests. The Friedman test demonstrated a statistically significant difference among the machine learning algorithms in credit-risk prediction, χ²(5) = 18.432, p = .002. Pairwise Wilcoxon tests showed that XGBoost significantly outperformed random forest (Z = -2.147, p = .032), neural network (Z = -1.983, p = .047), SVM (Z = -3.412, p = .001), decision tree (Z = -4.236, p = .001), and logistic regression (Z = -4.891, p = .001). The difference between random forest and neural network was not statistically significant (Z = -1.724, p = .085), whereas random forest significantly outperformed SVM (Z = -2.638, p = .008). The findings indicate that machine learning, particularly XGBoost and hybrid learning architectures, can improve customer experience in banking by supporting more accurate, reliable, explainable, and less disruptive service decisions, although continuous monitoring and retraining remain necessary under changing operational conditions.