syifbhuiyan

Algorithmic Fairness & Information Asymmetry in Digital Finance

Theme: Human-Centered AI / Information Ethics / Fintech

Goal

To audit a credit scoring model for demographic bias and mitigate information asymmetry using “Fairness, Accountability, and Transparency” (FAccT) principles.

Tools Used

Python (Scikit-Learn, Pandas)
Fairlearn (Microsoft’s bias audit toolkit)
Matplotlib (Data Visualization)

Project Overview

Coming from an agri-fintech background, I investigated how alternative data sources in credit scoring can introduce unintended bias against specific demographics. This project simulates a credit risk scenario using the Home Credit Default Risk dataset.

The core research question:
How do automated decision systems penalize marginalized groups, and can we fix it without breaking the model?

Key Findings & Insights

Bias Detected:
The baseline Logistic Regression model exhibited severe demographic disparity. It predicted “High Risk” for 90.5% of Male applicants compared to only 14.0% of Female applicants.
Bias Mitigated:
By applying a ThresholdOptimizer with Demographic Parity constraints, I successfully recalibrated the model.
Outcome:
The selection rate was equalized to approximately 41.6% for both groups, ensuring the algorithm treats applicants based on financial history rather than gender.

Visualizing the Impact

The chart below demonstrates the dramatic shift in selection rates before and after applying the fairness optimization:

Fairness Audit Chart showing balanced selection rates

Project Files & Code

This project includes a detailed Research Note and the full reproducible Python code.