Mitigating Algorithmic Bias in Predictive Models

Аннотация

This article considers the issue of systematic errors in predictive machine-learning models generating disparate outcomes for different social groups and OPEN ACCESS proposes a holistic approach to its mitigation. The risks SUBMITED 21 March 2025 and increasing legal requirements, along with corporate ACCEPTED 24 April 2025 PUBLISHED 29 May 2025 commitments to ethical AIs, drive the relevance of this VOLUME Vol.07 Issue 05 2025 study. The work herewith attempts to develop a bias- source taxonomy at data collection and annotation, CITATION Tamanno Maripova. (2025). Mitigating Algorithmic Bias in Predictive proxy-feature selection, model training, and Models. The American Journal of Engineering and Technology, 7(05), 192– deployment stages; also, it tries to compare pre-, in-, 201. https://doi.org/10.37547/tajet/Volume07Issue05-19. and post-processing methods' effectiveness on COPYRIGHT representative datasets measured by demographic © 2025 Original content from this work may be used under the terms of the creative commons attributes 4.0 License. parity, equalized error rates, and disparate impact. This article is unprecedented in undertaking a two-level approach: first, a systematic review of regulatory definitions (NIST, IBM) and case studies (COMPAS, healthcare-service prediction, face recognition) that identified key bias factors from sample imbalance to feedback loops; second, an empirical comparison of Reweighing, adversarial debiasing, threshold post- processing techniques alongside flexible multi-objective strategies—YODO (via AI Fairness 360 and Fairlearn libraries)—considering acceptable accuracy losses. The root source of unfairness remains data bias; hence, pre- processing must be undertaken (rebalancing, synthetic oversampling), while in- and post-processing can essentially harmonize group metrics at some cost in accuracy reduction Furthermore, without continuous online monitoring and documentation (datasheets, model cards), the balanced model risks losing fairness due to dynamic feedback effects. Bringing together technical fixes with rules and making the audit process official ensures the ability to copy and openness, which is key for long-term faith in AI systems. This article will help machine-learning builders, AI-responsibility experts, and checkers find ways to find, gauge, and lessen algorithmic bias in live models. The American Journal of Engineering and Technology 192 https://www.theamericanjournals.com/index.php/tajet Keywords: algorithmic bias, fairness, pre-processing, in- systematic errors that lead to unfair outcomes [1, 2] and processing, post-processing, demographic parity, an empirical analysis of the COMPAS tool demonstrating equalized odds, disparate impact, AI Act, NIST AI RMF, real cases of discrimination in judicial predictions [4]. To model cards. detect data biases, we analyzed model performance across groups. Specifically, we compared top-5 Introduction: Algorithmic bias refers to the systematic classification accuracy on ImageNet for images from error of running a machine-learning model in such a way regions with different income levels [5] and examined that causes members of different social groups to gender-recognition errors in commercial systems across receive drastically different predictions or decisions. It “race–gender” combinations [6]. develops when a model takes on historic inequities in the data, amplifies them through its training process, or To mitigate bias, three classes of technical strategies applies them to new situations where implicit were considered. The first line of defense comprises pre- correlations stand in for causal connections. processing methods, such as Reweighing, that adjust the Consequently, some groups are perceived as “risk- weights of training-set instances without altering the neutral” by default and others as “high risk” a priori, algorithm, achieving a disparate impact of 1.0 on the though in reality, event probabilities are equal. This is Adult dataset [18, 19]. The second class includes in- precisely how the National Institute of Standards and processing techniques that embed fairness constraints Technology (NIST) describes the problem—as a directly into the loss function: adversarial debiasing consequence of quantitative methods “flattening” rich achieved equalized odds parity with no more than a 2% social context into numerical categories, creating an reduction in overall accuracy [8]. The third line entails illusion of objectivity—while IBM defines it as post-processing algorithms that adjust model output “systematic errors that produce unfair outcomes” [1, 2]. probabilities via threshold optimization to balance group error rates [20]. Concurrent legal, reputational, and social effects confirm the significance of this issue for business and The legal and regulatory justification of the approach is society. The legal risk is evident: under the EU AI Act, ensured by mapping these technical practices to the violations of the pr