Data Quality and Fairness in ML

Research on improving fairness and robustness in machine learning models through data quality management

UCSD Halıcıoğlu Data Science Institute (Oct 2023 - Oct 2024)
Research Engineer | Supervisor: Babak Salimi | La Jolla, CA

This project focuses on developing a framework to manage and improve data quality, with the goal of enhancing fairness and robustness in machine learning models. The work involves creating tools to handle data biases and conducting experiments to assess the effectiveness of these tools.

DOI: Stress-Testing ML Pipelines with Adversarial Data Corruption
Code: https://github.com/lodino/savage

  • Built the error-injection pipeline of a black-box ML stress tester (beam search + Optuna Bayesian optimization).
  • Benchmarked 8 data-cleaning, 4 debiasing, and 2 uncertainty-quantification methods; at 10% corruption the framework cut AUC to 0.58 vs. 0.79–0.85 for random corruption and gradient-based poisoning.

References