Mitigating Bias in Machine Learning: A Comprehensive Review and Novel Approaches

Mahdi KhaliliAAAI 2025 · DOI

Introduction

As machine learning (ML) systems are increasingly deployed in high-stakes environments—including recruitment, financial lending, and criminal justice—the risk of "automated discrimination" has become a primary concern for researchers and practitioners alike. Bias can permeate the ML pipeline at multiple stages, originating from skewed historical data, biased objective functions, or unfair model outputs. This paper addresses the critical gap between theoretical fairness definitions and the practical implementation of these constraints, specifically focusing on how to mitigate bias without significantly compromising model accuracy.

A Taxonomy of Fairness Interventions

The research establishes a structured taxonomy of fairness, categorizing interventions into a three-part framework: Pre-processing, In-processing, and Post-processing. Within this framework, the author introduces three novel strategies designed to combat algorithmic bias effectively.

1. Generative AI for Pre-processing

Traditional pre-processing often involves filtering or re-weighting existing datasets, which can lead to information loss. This paper proposes a shift toward Fair Data Synthesis. By modifying generative models to incorporate fairness constraints, researchers can produce synthetic datasets that are inherently more equitable. When downstream tasks are trained on this "hallucinated" fair data, they inherit these equitable properties, bypassing the limitations of biased historical records.

2. Global Optimization for In-processing

Fair learning often requires solving complex, non-convex optimization problems. Historically, these have been avoided or simplified because they are prone to getting trapped in local minima, leading to sub-optimal fairness-accuracy trade-offs. The author presents a mathematical framing that allows these non-convex problems to be solved for global optimality. This ensures that the model achieves the best possible performance within the required fairness constraints, rather than settling for a local approximation.

3. SVD-based Post-processing

For models that are already trained, the paper introduces a "surgical" approach to de-biasing using Singular Value Decomposition (SVD). This method identifies specific mathematical components (singular vectors) within the model that contribute to unfair decisions. By isolating and removing these specific components, the model can be de-biased post-hoc. This targeted extraction allows for the removal of discriminatory patterns while preserving the model’s overall predictive utility and accuracy.

Key Insights and Breakthroughs

The research highlights several qualitative breakthroughs in the field of algorithmic fairness:

  • Fairness as Synthesis: Rather than merely fixing broken data, generative AI allows for the creation of a fairer version of reality for training purposes.
  • Overcoming Non-convexity: The work demonstrates that the mathematical complexity of fair ML is not a dead end; global optimality is achievable with the right formulation.
  • Surgical Precision: Bias is often localized within specific model components. Using SVD, these components can be extracted without destroying the model's functionality.
  • Individual vs. Group Fairness: The paper reinforces the distinction between individual fairness (treating similar individuals similarly) and group fairness (equalizing outcomes across demographics), providing a toolkit that addresses both through different stages of the ML pipeline.

Conclusion

By providing a robust taxonomy and three distinct mathematical approaches to bias mitigation, this work moves the field closer to practical, fair-by-design machine learning. Whether through synthetic data generation, global optimization, or surgical post-processing, these methods offer a path forward for deploying equitable AI in sensitive, real-world applications.

Summary updated

BibTeX
@inproceedings{khalili2025mitigating,
  author = {Mahdi Khalili},
  title = {{Mitigating Bias in Machine Learning: A Comprehensive Review and Novel Approaches}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i27.35107},
}
Figure