FairLens — Interpretable Bias Detection

---

FairLens is a fairness-aware ML demo that combines:

• Bias Detection — Quantify disparities in accuracy, false positive/negative rates across groups
• Bias Mitigation — Apply methods like Reweighting and Exponentiated Gradient (Fairlearn)
• Interpretability — Use SHAP explainability to visualize feature importance

Built as a multi-agent pipeline with a polished Streamlit UI, it highlights responsible AI practices for healthcare and income prediction datasets.

---

✨ Features

• Multi-agent pipeline: modular agents for data loading, modeling, mitigation, evaluation, explanation.
• Datasets supported:
  • Pima Indians Diabetes (health prediction)
  • Adult Income (census data)
• Bias mitigation methods:
  • Baseline (no mitigation)
  • Reweighting (sample-weight rebalancing)
  • ExponentiatedGradient (Fairlearn reductions)
• Explainability: SHAP value plots for feature importance.
• Fairness metrics: accuracy, precision, recall, FPR/FNR by group + fairness gaps.
• Interactive dashboard: Streamlit UI with tabs, cards, and comparison tables.
• Static report: auto-generated report.ipynb with reproducible analysis.

---

🌐 Demo (Streamlit)

🚀 Live Demo

👉 Try the Streamlit app here

• Dataset toggle: Pima Diabetes (healthcare) and Adult Income (socioeconomic)
• Compare Baseline vs Mitigation side-by-side
• Visualize SHAP explanations and fairness gaps (e.g., FPR difference, accuracy difference)

---

📊 Example Results

Overall metrics (Baseline vs Reweight vs ExpGrad):

Metric	Baseline	Reweight	ExpGrad
Accuracy	0.74	0.74	0.75
Precision	0.67	0.67	0.71
Recall	0.50	0.50	0.49

Fairness gap (FPR difference): reduced after mitigation.

---

📂 Project Structure

📂 Repo Structure FairLens/ │ ├── pmas/ # Core multi-agent system │ ├── agents/ # Modular agents (data, model, explain, mitigate) │ ├── orchestrator.py # Orchestrator │ └── main.py # Pipeline entrypoint │ ├── webui/ # Streamlit dashboard │ └── streamlit_app.py │ ├── outputs/ # Generated SHAP plots, metrics, CSVs ├── tools/ # Reporting utilities │ └── generate_report.py # Creates report.ipynb │ ├── assets/ # Screenshots for README │ ├── streamlit_dashboard.png │ └── shap_example.png │ ├── report.ipynb # Reproducible analysis notebook ├── requirements.txt └── README.md

---

🛠️ Tech Stack

• Python 3.12
• scikit-learn — ML models
• Fairlearn — Fairness metrics & reductions
• SHAP — Interpretability
• Streamlit — UI

---

📘 Report

A reproducible notebook report.ipynb is included with:

• End-to-end pipeline runs
• SHAP plots inline
• Fairness gap calculations

---

✨ Why This Project?

FairLens demonstrates responsible AI deployment:

• Detecting and explaining bias
• Applying mitigation techniques
• Providing a transparent, interactive UI

This is especially relevant for AI in healthcare and socioeconomic decision-making.

---

📜 License MIT License. Free to use and adapt.

👩‍💻 Built by Samriddhi Sharma — fairness, interpretability, and ML systems.