Freeze! Is This Model Fair?

Technology(s) Used: Scikit-learn, Fairness Indicators, Platt Scaling, Random Forest, AdaBoost, Python

• Objective: Analyzed over two decades of NYPD stop-and-frisk data to detect and mitigate predictive bias in arrest outcomes across sensitive demographics (race, gender, and age).
• Fairness Interventions: Implemented a three-stage fairness pipeline, applying pre-processing techniques such as uniform sampling and preferential sampling to mitigate dataset bias, and post-processing methods including Platt scaling to adjust predictions and enforce equalized odds after model training.
• Modeling: Built robust classifiers using in-processing strategies, training Logistic Regression, Random Forest, and AdaBoost models, and assessed performance using balanced accuracy and demographic parity metrics.
• Results: The post-processing + Random Forest pipeline emerged as the best performer, increasing balanced accuracy from 0.68 to 0.71 while achieving significantly more equitable TPR/FPR across groups.