Fairness-Aware Credit Scoring

A Comparative Study and a Hybrid Model Approach

Machine learning models used in credit scoring often inherit historical bias from financial data, leading to discriminatory outcomes against protected demographic groups. This project evaluates 12 fairness-aware ML models across four real-world credit datasets, introduces a novel composite metric called the Responsibility Score (RS), and develops a hybrid fairness pipeline applied to the HMDA dataset.

---

Research Questions

• RQ1 — How do fairness-aware ML models perform across multiple real-world credit datasets in terms of accuracy and fairness?
• RQ2 — Can a composite metric (Responsibility Score) better capture the balance between fairness and accuracy?
• RQ3 — Which fairness strategy — pre-, in-, or post-processing — offers the most consistent and responsible outcomes?
• RQ4 — Can a hybrid model combining multiple fairness strategies improve outcomes further?

---

Datasets

Dataset	Source	Protected Attribute	Task
German Credit	UCI ML Repository	Age	Good/bad credit risk
Give Me Some Credit	Kaggle	Age	Loan default prediction
Taiwan Credit Default	UCI ML Repository	Age	Credit card default
Credit Card Approval	Kaggle	Age	Loan approval
HMDA 2018 (Illinois)	CFPB	Race, Ethnicity, Age	Mortgage approval (hybrid model)

---

Models Evaluated

Baseline

• XGBoost

Pre-processing

• Learning Fair Representations (LFR)
• Disparate Impact Remover
• Reweighing

In-processing

• Exponentiated Gradient Reduction
• Grid Search Reduction
• GerryFair Classifier
• Prejudice Remover
• Adversarial Debiasing

Post-processing

• Equalized Odds
• Calibrated Equalized Odds
• Reject Option Classification

---

Responsibility Score (RS)

We introduce RS as a single interpretable metric combining accuracy and fairness:

RS = α × Accuracy − β × |SPD| − γ × |EOD|

With weights α = 1, β = 2, γ = 2 to penalize fairness violations more than prediction error.

Top models by RS across all datasets:

Rank	Model	Avg RS
1	LFR	0.912
2	Equalized Odds	0.775
3	Grid Search	0.771
4	Disparate Impact Remover	0.742
5	Exponentiated Gradient	0.733

---

Key Findings

• Pre-processing methods (especially LFR and Reweighing) achieved the highest average RS with the least variance across datasets — the most consistent strategy.
• LFR was the top-performing model overall: high accuracy with low SPD and EOD.
• Post-processing methods were least stable and often reduced accuracy.
• Reject Option Classification produced Disparate Impact values exceeding legal thresholds.
• The hybrid pipeline (DIR + EGR) on HMDA achieved 95.9% accuracy with DI ≈ 0.94 and SPD ≈ −0.047 — the best fairness-accuracy trade-off.
• SHAP analysis confirmed the hybrid model relied on financial reasoning (loan amount, DTI, property value) rather than demographic proxies.
• Counterfactual fairness (WCF Score = 0.9249) confirmed high individual-level fairness.

---

Dependencies

Install required libraries:

pip install aif360[all]
pip install fairlearn
pip install xgboost
pip install scikit-learn pandas numpy matplotlib shap

Note: Notebooks were developed in Google Colab. If running locally, remove the drive.mount() cells and update any file paths accordingly.

---

Fairness Metrics Used

Metric	Description
Accuracy	Standard predictive accuracy
Statistical Parity Difference (SPD)	Difference in positive prediction rates between groups
Equal Opportunity Difference (EOD)	Difference in true positive rates between groups
Disparate Impact (DI)	Ratio of positive outcomes between unprivileged and privileged groups (ideal ≈ 1.0)
Responsibility Score (RS)	Novel composite metric: accuracy penalized for fairness violations

---

Paper

Full write-up available in Fairness-Aware_Credit_Scoring.pdf

Citation:

Niyati Malik and Niket Pathak. 2025. Fairness-Aware Credit Scoring:
A Comparative Study and a Hybrid Model Approach.
University of Illinois Chicago. CS516.

---

Authors

Niket Pathak — npath5@uic.edu
Niyati Malik — nmali@uic.edu
University of Illinois Chicago