Introduction
Artificial Intelligence in Education (AIEd) systems are increasingly deployed to automate high-stakes decisions, such as resource allocation and scholarship awards. While these systems aim for efficiency, they often inherit and amplify existing societal biases. Traditional fairness-aware algorithms typically treat fairness as a static constraint at a single point in time. However, in real-world educational settings, decisions are sequential. The authors of this paper identify a critical phenomenon: Fairness Instability. When algorithms attempt to correct for long-term bias without considering temporal dynamics, they often resort to "drastic mitigation"—erratic policy shifts that over-correct for past imbalances, leading to unpredictable and ethically questionable outcomes for student cohorts.
Methodology: The FAiRDAS Framework
To address this, the researchers propose FAiRDAS (Fairness-Aware Ranking as Dynamic Abstract System). This framework treats the evolution of fairness and accuracy metrics as a linear discrete dynamic system. Instead of viewing fairness as a fixed destination, FAiRDAS defines a "target trajectory" that smoothly guides the system toward fairness goals over time.
The system operates by applying "action vectors" ($\theta$)—penalties or rewards based on sensitive attributes like socioeconomic status—to predicted student scores. By using Sequential Least Squares Programming (SLSQP), FAiRDAS optimizes these actions to ensure that the resulting metrics follow a stable, predictable path defined by a convergence factor ($\lambda$). This approach prevents the "flickering" behavior seen in standard greedy optimizers.
Experimental Evaluation
The framework was evaluated using a longitudinal dataset from the Canarian Agency for Quality Assessment and Accreditation (ACCUEE), spanning 2015 to 2019. The model predicts student math performance, with the Economic, Social, and Cultural Status (ESCS) index serving as the protected variable.
The researchers simulated 100 sequential batches of students to observe how different ranking policies evolved. They compared FAiRDAS against a greedy baseline that minimized unfairness at each step without considering temporal stability.
Key Results and Discussion
The experiments demonstrated that FAiRDAS successfully achieves "smooth fairness" without sacrificing performance:
- Stability vs. Volatility: While the greedy baseline exhibited high instability in its mitigation actions ($m_{Actions} \approx 0.33$), frequently jumping between extreme penalties, FAiRDAS maintained a significantly smoother profile ($m_{Actions} \approx 0.044$).
- Performance Parity: FAiRDAS achieved comparable levels of fairness (via the Disparate Impact Discrimination Index) and accuracy (Sum of Absolute Errors) to the baseline. It reached the same ethical goals but through a more justifiable and stable policy path.
- Interpretability: By stabilizing the resulting metrics, the system naturally produces stable actions. This consistency is vital for human stakeholders—educators and policymakers—who require predictable systems to maintain public trust.
Conclusion
FAiRDAS represents a shift in algorithmic fairness from static snapshots to dynamic trajectories. By modeling fairness as a controlled path, it avoids the pitfalls of drastic policy swings in educational resource allocation. Future work will explore the application of this framework to other domains like healthcare and lending, as well as more complex action spaces beyond simple score adjustments.
Summary updated