Detecting Hidden Bias in Machine Learning Models: A Systematic Approach to Fairness Evaluation and Mitigation

Author: Debabrata Pruseth
Publication Date: 2025/11/08
Document Type: Technical Note / Research Article
Language: English

Abstract

Machine learning models increasingly influence high-stakes decisions in employment, lending, education, healthcare, insurance, and public services. Although these systems are often evaluated using aggregate accuracy, precision, recall, or area-under-curve metrics, such measurements can conceal unequal performance across demographic groups. A model may appear statistically effective overall while producing systematically unfavorable outcomes for a protected or underrepresented group. This study presents a systematic applied framework for detecting hidden bias in machine learning models and translating fairness evaluation into practical mitigation steps. The methodology begins with identification of decision context, target outcome, protected attributes, and possible proxy variables. It then evaluates model outcomes using group-level selection rates, error rates, false positive rates, false negative rates, and disparate impact ratios. The framework emphasizes that fairness cannot be reduced to a single metric; rather, fairness evaluation requires contextual interpretation, stakeholder review, domain-specific thresholds, and continuous monitoring. Mitigation strategies are organized into pre-processing, in-processing, and post-processing interventions, including dataset rebalancing, feature review, fairness-constrained learning, threshold adjustment, and human oversight. The study also discusses responsible AI governance, including transparency, auditability, documentation, privacy protection, and risk of overclaiming fairness. The contribution of this work is a practical, structured fairness-evaluation workflow that supports both technical analysis and governance-oriented decision-making. The study is conceptual and implementation-oriented; it does not claim benchmark superiority or empirical validation across datasets.

Keywords
Machine learning fairness; hidden bias; responsible AI; algorithmic bias; disparate impact ratio; demographic parity; equalized odds; fairness evaluation; bias mitigation; model governance; AI audit; Fairlearn; AI Fairness 360; explainable AI; trustworthy AI

Download Research PDF

Detecting Hidden Bias in Machine Learning Models: A Systematic Approach to Fairness Evaluation and Mitigation

Suggested Citation
Pruseth, D. (2026). Detecting Hidden Bias in Machine Learning Models: A Systematic Approach to Fairness Evaluation and Mitigation. Debabrata Pruseth AI blog.

Companion Note
This page provides the abstract and full-text PDF for the research version of the article. A companion blog post explains the same work in a more narrative and implementation-focused style.

Read the companion blog:
https://debabratapruseth.com/how-to-detect-hidden-bias-in-your-ml-model-a-step-by-step-tutorial/