A Systematic Review and Taxonomy of Bias Detection in Machine Learning for Scholarship Allocation in Educational Decision Systems

The rapid adoption of Machine Learning (ML) in educational decision-making, particularly scholarship allocation, has raised critical concerns about fairness, bias propagation, and institutional accountability. However, existing research on how bias is detected, measured, and mitigated in these systems remains fragmented across methods, domains, and evaluation practices. This study conducts a PRISMA-guided systematic review to explore machine learning-driven bias detection approaches in education and scholarship, integrating findings from Scopus, Web of Science, IEEE Xplore, and the ACM Digital Library. The synthesis examines bias variants, fairness metrics, methodological trends, and evaluation practices across selected studies. Findings show that representation and measurement bias dominate the literature, while label and deployment biases are less explored. Statistical and group fairness metrics are most commonly used, whereas causal inference and in-model fairness approaches remain underdeveloped. Major methodological limitations identified include the scarcity of high-quality datasets, inconsistent reporting practices, limited reproducibility, and inadequate evaluation of fairness across the entire model lifecycle. To address these issues, the study proposes a domain-specific taxonomy for scholarship allocation that structures fairness analysis across bias source, detection stage, method type, fairness metric, and educational context. The framework consolidates fragmented evidence and highlights research gaps in causal fairness, deployment monitoring, and longitudinal bias analysis.