Ethically Sound Deep Learning Model for Credit Card Offer Acceptance Prediction

📌 Overview

This project develops and evaluates a deep learning model to predict whether a customer will accept a credit card offer, while placing a strong emphasis on ethical AI practices.

We address potential biases in the dataset, especially those related to income-based discrimination, and demonstrate how to create fair, explainable, and trustworthy AI systems using modern tools like SHAP, Fairlearn, and AIF360.

---

📊 Dataset

• Source: Kaggle Credit Card Offer Dataset
• Features: Income Level, Credit Rating, Average Balance, Mailer Type, Overdraft Protection, etc.
• Sensitive Attribute: Income Level

---

✅ Tasks Completed

1. Initial Deep Learning Model

   • 3-layer neural network
   • Trained using SMOTE-balanced data
   • Evaluated with F1 Score due to class imbalance

2. Explainability using SHAP

   • Visualizes feature impact for transparency
   • Shows model's decision-making logic

3. Bias Detection with Fairlearn

   • Detected unfair treatment across income groups

4. Ethical Redesign

   • Reweighing (Fairlearn)
   • Adversarial Debiasing (AIF360)
   • Compared fairness vs. performance

---

🧪 Results Summary

📊 Performance Metrics by Income Group

Model	Income Group	Accuracy	Precision	Recall	F1 Score	Selection Rate
Baseline	Low/Med Income	0.9005	0.0783	0.0523	0.0627	0.0425
	High Income	0.9264	0.0541	0.0606	0.0571	0.0412
Reweighed	Low/Med Income	0.8690	0.0748	0.0930	0.0829	0.0792
	High Income	0.8930	0.0435	0.0909	0.0588	0.0769
Adversarial	Low/Med Income	0.7281	0.0676	0.2558	0.1069	0.2408
	High Income	0.8874	0.0405	0.0909	0.0561	0.0825

---

🧠 Insights

• Baseline model favors accuracy but is unfair — large disparity in treatment and selection rates.
• Reweighed model slightly drops accuracy but brings excellent fairness without big performance hits.
• Adversarial Debiasing makes the strongest fairness push (especially in selection rate balance and recall for the disadvantaged group) — but sacrifices accuracy, especially for low/medium-income groups.

---

Metric	Original Model	Baseline	Fair (Reweighed)	Adversarial Debiasing
Accuracy	0.8322	0.8544	0.9350	0.8208
Precision	0.1373	0.0701	0.1081	0.0473
Recall	0.3474	0.1268	0.0195	0.1122
F1 Score	0.1968	0.0903	0.0331	0.0666
ROC AUC	—	0.5167	0.5015	0.5141
Statistical Parity Diff	0.0087	0.0023	-0.0129	0.1369
Disparate Impact	1.0769	0.9779	0.7904	123.7814
Equal Opp. Diff	-0.0107	0.0655	-0.0573	0.1221
Avg Odds Diff	-0.0005	—	-0.0344	0.1300
Theil Index	0.0871	—	0.0715	0.0840

---

⚖️ Key Ethical Takeaways

• Accuracy alone is not sufficient.
• Fairness-aware techniques like reweighing and adversarial debiasing can help mitigate bias.
• SHAP improves transparency and trust.
• Ethical trade-offs are necessary and application-dependent.

---

🛠️ Tools & Libraries

• Python, Pandas, Scikit-learn
• Keras + TensorFlow (Deep Learning)
• SHAP (Explainability)
• Fairlearn, AIF360 (Fairness)
• SMOTE (Balancing)

---

▶️ How to Run

1. Click the badge at the top or open the Google Colab Notebook
2. Follow the cells step-by-step to preprocess data, train the models, and evaluate performance and fairness.

---

📄 Report

Report Link

---

🧑‍🎓 Author

Harshini V