What question did this study set out to answer?

The aim is to create a transparent and accurate machine learning ensemble model for predicting loan approvals while addressing fairness and bias.

June 26, 2026Open Access

An explainable ensemble learning framework for transparent and accurate loan approval prediction

Key Points

The aim is to create a transparent and accurate machine learning ensemble model for predicting loan approvals while addressing fairness and bias.
Developed an ensemble model using Random Forest and XGBoost in a soft-voting classifier.
Utilized SHAP and LIME for model transparency and interpretability, alongside a fairness audit on demographic attributes.
Benchmarked performance using Kaggle Loan Prediction dataset with 13,508 records and implemented 5-fold stratified cross-validation.
Achieved 95.16% accuracy and 0.92 F1-score, outperforming standalone models by 4–13 points.
Identified key decision factors: credit history, applicant income, and loan amount, aligned with credit-risk theory.
Revealed significant interaction effects between Loan Amount and Applicant Income, and Credit History and Property Area in decision-making.

Abstract

Highlights A machine learning-driven ensemble model is developed for efficient and automated loanapproval predictionThe ensemble model is benchmarked against conventional models using performance metricssuch as accuracy, precision, recall, and F1-score. Explainable AI (XAI) techniques including SHAP and LIME are integrated to enhancetransparency and interpretability of model predictions. Key factors influencing loan approval decisions are identified, promoting fairness and reducingbias in automated financial assessments. Abstract Loan approval is a high-volume, high-stakes decision that banks increasingly automate. Existing machine-learning systems are either accurate but opaque, or interpretable but inaccurate, and rarely report fairness diagnostics. This paper presents an integrated ensemble pipeline that combines Random Forest and XGBoost in a soft-voting classifier, jointly tuned by GridSearchCV with 5-fold stratified cross-validation, on the Kaggle Loan Prediction dataset (13, 508 records, 13 features). The pipeline includes reproducible preprocessing (median/mode imputation, label and one-hot encoding, StandardScaler), and is paired with a dual post-hoc explanation layer (global SHAP and local LIME with reported fidelity) and a group-fairness audit (demographic parity, equal opportunity, disparate impact) on the protected attributes Gender and Education. The ensemble achieved 95. 16% accuracy and 0. 92 F1-score at an 80: 20 split, outperforming standalone baselines by 4–13 points, with stable 10-fold cross-validation performance (0. 948 ± 0. 011). Credit history, applicant income, and loan amount were identified as the dominant decision drivers, consistent with established credit-risk theory. SHAP interaction analysis additionally surfaced a negative LoanAmount-ApplicantIncome interaction and a CreditHistory-PropertyAreaSemiurban interaction that are not visible in single-feature importance rankings.

Ask AI

Helpful

Bookmark

View Full Paper