Year: 2026
Type: article
Abstract: Student attrition in higher education is a multidimensional challenge shaped by demographic, academic, behavioral, and socioeconomic factors. This study proposes an AutoML 3.0 framework for early dropout prediction, grounded in educational persistence theory, learning analytics, and human-centered machine learning (HCML). The framework integrates evolutionary data-centric optimization (EDCA), domain-aware model selection, and human-in-the-loop oversight. Using a longitudinal dataset of 6,287 students (2014–2021), it embeds fairness auditing across age, gender, and socioeconomic subgroups, uncertainty calibration via Brier scores, and temporal cohort validation. The top configuration (FLAML + LightGBM) achieves 0.802 accuracy, 0.740 macro-F1 (95% CI: [0.71, 0.76]), and a Brier score of 0.14, reducing false negatives among older students by 18% and maintaining demographic parity below 5%. These results show that theoretically grounded, ethically aligned AutoML can deliver scalable and responsible early-warning systems in higher education.
Institutions: Pegaso University, University of Palermo
Language: English
Subfield: Computer Science Applications
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7225595242