This project builds a machine learning pipeline to predict loan defaults while ensuring fairness, transparency, and interpretability — aligning with ethical AI practices in finance. Beyond accuracy, the focus is on creating auditable and bias-aware credit risk models that financial institutions can trust.

Develop a binary classification model to predict whether a loan applicant is likely to default. The pipeline emphasizes:

• High recall (catching as many defaults as possible)
• Explainability (why the model predicted default)
• Fairness (ensuring no demographic group is unfairly penalized)

• Default cases were minority class (~15–20%), creating imbalance.
• Applied SMOTE (Synthetic Minority Oversampling Technique) to upsample the default class.
• Balanced dataset improved recall and generalization compared to baseline.

• Encoded categorical features (education, marital status, job type).
• Scaled continuous variables (income, loan amount, age).
• Handled missing values with median/mode imputation.
• Feature selection applied to retain only informative predictors.

• Trained Random Forest Classifier on SMOTE-balanced data.
• Compared baseline models with tuned Random Forest.
• Evaluated on three sets: training, test, and unseen test.

• Accuracy: 0.94
• Precision: 0.86
• Recall: 1.00
• F1-Score: 0.92
• ROC-AUC: 0.96

• Accuracy: 0.91
• Precision: 0.84
• Recall: 0.91
• F1-Score: 0.87
• ROC-AUC: 0.91

• Accuracy: 0.89
• Precision: 0.54
• Recall: 0.77
• F1-Score: 0.64
• ROC-AUC: 0.84

👉 Insight: Strong generalization on test data, but precision dropped on unseen data — showing real-world challenges. This shifted the project’s emphasis from pure accuracy to explainability and fairness.

Applied post-hoc explainability methods to build trust:

• SHAP: Global feature importance & local force plots
  → Income, job tenure, car ownership were top predictors.
• LIME: Local explanations for individual applicants.
• DiCE: Counterfactuals showing actionable “what-if” scenarios (e.g., higher income or stable employment reduces default risk).
• Anchors (Alibi): If-then rules for transparent decision boundaries.

• Used AIF360 and Fairlearn to audit bias.
• Metrics:
  • Demographic Parity Difference
  • Equalized Odds Difference
• Bias detected across gender and marital status.
• Bias Mitigation: Retraining with fairness constraints reduced disparate impact by ~12%.

• SHAP summary plots: Showed top drivers of default risk.
• ALE & PDP plots: Illustrated how income and tenure influence probability of default.
• Fairness comparison plots: Showed group-level disparities before and after mitigation.

• Built a credit risk model achieving 0.91 accuracy & 0.91 ROC-AUC on test data.
• On unseen data, recall remained solid (0.77) but precision dropped, highlighting need for robust interpretability.
• Delivered transparent explanations for individual applicants, enabling human loan officers to audit predictions.
• Reduced demographic bias by 12% after retraining, making the model more compliant with responsible AI guidelines.
• Created actionable insights for financial institutions to evaluate creditworthiness with fairness and accountability.

Loan_Default_Interpretability/

• Loan_Default_Interpretability.ipynb → Full notebook with preprocessing, modeling, explainability
• README.md → Documentation (this file)
• Training Data.csv → Sample dataset
• assets/ → Plots (SHAP, LIME, fairness audit, PDP)

• Python: pandas, scikit-learn, imbalanced-learn
• ML Models: Random Forest, EBM (Explainable Boosting Machine)
• Explainability: SHAP, LIME, DiCE, Anchors, InterpretML
• Fairness Auditing: AIF360, Fairlearn
• Visualization: Matplotlib, Seaborn, PDPBox, PyALE
• Dashboarding: Tableau (for stakeholder-facing summaries)

• Deploy as a Streamlit app for real-time loan risk scoring.
• Explore ensemble fairness-aware methods (e.g., adversarial debiasing).
• Integrate external credit bureau data for stronger predictions.