FairLoans– Debiasing Loan Approval Models for Responsible AI

Builds a fair and explainable AI system to audit, detect, and mitigate bias in loan approval decisions using XGBoost, SHAP, and Fairlearn. Promotes responsible AI through transparency and fairness.

Comment

Inspiration

Bias in algorithmic loan approvals can silently harm underrepresented groups, especially in financial services. We wanted to build a responsible AI system that not only detects bias but actively reduces it — with full transparency and real-time explainability.

What it does

FairLoans is an AI-powered auditing and mitigation pipeline that:

  • Detects bias in historical loan approval data using Fairlearn
  • Applies Exponentiated Gradient mitigation with Demographic Parity constraints
  • Visualizes fairness metrics like demographic parity difference, equalized odds, and selection rate
  • Provides SHAP explainability to demystify model behavior
  • Deploys a Streamlit Dashboard for interactive simulation, fairness auditing, and test predictions

How we built it

  • Exploratory analysis using Pandas, Seaborn, and Matplotlib
  • Fairness detection & mitigation with Fairlearn
  • Model training using XGBoost and fairness-constrained algorithms
  • Interpretability via SHAP values and summary plots
  • Frontend dashboard in Streamlit with upload, visualize, and predict modules
  • Submission pipeline for real-world test set predictions

Challenges we ran into

  • Handling fairness–accuracy trade-offs in real-world imbalanced datasets
  • Encoding categorical variables consistently across test/train for reliable results
  • Ensuring transparency without sacrificing performance
  • Explaining decisions clearly for non-technical users

Accomplishments that we're proud of

  • Successfully detected and quantified bias in a real-world loan approval dataset
  • Implemented fairness mitigation using Exponentiated Gradient with Demographic Parity constraints
  • Reduced Equalized Odds Difference from 0.17 to 0.09 without severely sacrificing model accuracy
  • Developed a fully functional Streamlit Dashboard that supports:
    • Interactive fairness audits
    • SHAP-based explainability
    • Real-time loan approval simulation
    • Final test-set predictions
  • Delivered a clean and reproducible codebase with modular scripts for training, explanation, and prediction
  • Built a submission-ready CSV file for Devpost that reflects ethical and accountable machine learning
  • Deployed the entire solution live using Streamlit Cloud, making it publicly accessible
  • Learned to balance performance, fairness, and interpretability — the trifecta of responsible AI

What we learned

  • Fairness is not just a metric — it’s a design choice that affects real lives
  • SHAP can be powerful in making AI models explainable and auditable
  • Tools like Fairlearn and AIF360 can help bridge ethical AI and engineering

What's next for FairLoans– Debiasing Loan Approval Models for Responsible AI

  • Extend FairLoans to other sensitive domains (e.g., hiring, healthcare)
  • Build an enterprise-grade SaaS dashboard for fairness audits
  • Enable continuous monitoring of model drift and fairness over time

Built With

Share this project:

Updates