Bias Detection and Fairness Improvement in Machine Learning Models

Abstract

The growing use of machine learning models in sensitive areas of choice has come under great concern due to the issue of algorithm bias and fairness. The inputs to ma-chine learning systems are frequently based on historical data that brings the current inequalities existing in society and may result in discriminatory behaviour against the groups that are being defended against (gender, race, ethnicity, or nationality). The study explores the theme of bias detection and fairness enhancement in machine learn-ing models through the synthesis of results of empirical studies, systematic reviews, and application-specific work. It looks into the main origins of bias, such as data-based, algorithmic, and human-based bias, and explores the most commonly used measures of fairness, such as statistical parity, disparate impact, equalised odds, and equal opportunity. The study also examines bias detection methods and mitigation approaches realised in pre-processing, in-processing and post-processing phases of the machine learning pipeline. Past studies are discussed with regard to their experimental evidence to illustrate the trade-offs of predictive performance and fairness, with fair-ness-aware methods having significant ability to lessen bias at acceptable levels of accuracy. As well, the study highlights the necessity of transparency, explainability, and ethical considerability confined in the production of reliable artificial intelligence systems. Additionally, the research shows that it is vital to have context-sensitive evaluation systems alongside combined equity targets that can facilitate responsible and just implementations of machine learning apprehensions in various real-world sectors.