Document Type : Research Paper

Authors

• Mohammad Fathian
• Mohammadreza Rasouli
• Sahar Oladi

Iran University of Science and Technology

Abstract

Ensuring equity in educational assessment is essential for providing equitable learning opportunities to all students, regardless of their gender, race, or socio-economic background. However, persistent disparities in large-scale educational evaluations indicate that data-driven models may unintentionally amplify existing inequities when trained on biased data. This study evaluates the effectiveness of two causal fairness–modification algorithms—MData and Matrix Factorization (MF)—in mitigating discrimination within the TIMSS student-achievement dataset. MData applies threshold-based causal label modifications, whereas MF reconstructs group-level distributions by enforcing conditional independence through factorization. Experimental results demonstrate that both algorithms substantially mitigate discrimination across gender, race, and socio-economic status. MData consistently preserves superior predictive accuracy, while MF achieves greater fairness gains under an appropriate configuration. Together, these findings underscore the potential of causal pre-processing techniques to modify biased datasets and enhance equity in educational analytics. They also highlight the importance of integrating fairness-aware data modification into large-scale assessment pipelines.

Keywords

• Educational Fairness
• Causal Learning
• TIMSS dataset
• Modification
• Pre-processing Algorithms