The Illusion of Fairness: A Systematic Study of Unsupervised Domain Adaptation on Sparse Educational Data

Through a systematic evaluation using the Open University Learning Analytics Dataset (OULAD), this study examines Correlation Alignment (CORAL) across 42 transfer pairs (N = 32,593). The findings reveal a pervasive failure mode: CORAL consistently degraded model accuracy in 93% of cases (Mean Drop:-18.0%) and failed to yield reliable improvements in fairness metrics. These results suggest that the "feature structure destruction" inherent in sparse, one-hot encoded data, coupled with significant label shift, renders naive alignment methods counterproductive in educational contexts.