Differential Privacy and Fairness on COMPAS

Overview

This project from MAT180 investigates two key themes in trustworthy ML:

• Differential Privacy (DP): How injecting calibrated noise impacts K-Means clustering objective and cluster quality.
• Algorithmic Fairness: Diagnosing group disparities in recidivism predictions using COMPAS risk scores.

Goals

• Implement differentially private K-Means variants with Laplace/Gaussian noise.
• Quantify fairness with Equalized Odds and Sufficiency across protected groups in COMPAS.
• Summarize trade-offs and qualitative patterns in a short writeup and poster.

Data

• COMPAS subset: /projects/dp-fairness/data/compas-scores.csv
• Synthetic/teaching data: /projects/dp-fairness/data/kmeansexample.asc

Methods

• DP K-Means: add noise to counts/centroid updates and study objective/cluster assignments vs. baseline.
• Fairness metrics: compute confusion matrices by group; compare error rates and calibration-like statistics.

Deliverables

• Report: /projects/dp-fairness/reports/dp_fairness_report.pdf
• Poster: /projects/dp-fairness/reports/machine_learning_fairness_poster.pdf
• Notebooks:
  • DP K-Means: /projects/dp-fairness/notebooks/differential-privacy.ipynb
  • Fairness metrics: /projects/dp-fairness/notebooks/fairness.ipynb

Outcomes

• DP noise degrades K-Means objective and cluster stability as privacy increases; moderate noise retains reasonable structure on simple datasets.
• COMPAS analysis highlights group-level differences in error rates and score–outcome relationships, consistent with known fairness concerns.

These artifacts are designed for teaching: the notebooks are concise, with code and outputs that can be reused in coursework or demos.