| We examine whether pool-based active learning can improve predictive performance without worsening group fairness under scarce labels. We compare a fairness-only baseline with three lightweight query strategies on two gender-sensitive learning-analytics tasks: MBA admissions and OULAD student success prediction. The strategies are random sampling, uncertainty entropy, and uncertainty margin. The pipeline evaluates decision trees and random forests under label budgets from 20% to 90%, starting from a 10% labeled seed set and applying reweighing before final supervised training. RF+UE achieves the highest accuracy in six of eight dataset-budget settings and in all OULAD settings, whereas the DT+FO baseline achieves the lowest equalized-odds difference in seven of eight settings. Averaged across budgets, RF+UE reaches accuracy 0.838 on MBA and 0.790 on OULAD, while DT+FO reduces mean equalized-odds difference to 0.024 and 0.018. Active learning does not solve fairness automatically; instead, it shifts the operating point along a measurable accuracy–fairness frontier. For learning-analytics deployment, lightweight active learning is most compelling when predictive utility is primary, whereas fairness-first reweighted baselines are safer when disparity reduction is the dominant requirement |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.