Territorial fairness in large-scale academic risk prediction: comparing national and state-level machine learning models in Brazil

Conference item

Early identification of students at academic risk is a central challenge for large and decentralized educational systems. In countries such as Brazil, pronounced regional disparities raise concerns not only about predictive performance, but also about whether machine learning models generalize equitably across territories. This study examines the role of regionalization in academic risk prediction by comparing national and state-level models trained under a unified experimental pipeline using longitudinal administrative data from over six million upper secondary student enrollments across all Brazilian states. Multiple supervised learning algorithms are evaluated, with Random Forest selected for detailed analysis due to its robust overall performance. Territorial fairness is assessed through an operationalization of Equal Opportunity and Equalized Odds based on state-level true and false positive rates. Results show that while national and state-level models achieve similar aggregate performance, substantial disparities persist in Recall across states. State-level models improve local risk detection in a small subset of states, often at the cost of increased false positives. These findings indicate that regional specialization is not uniformly beneficial and should be understood as a context-dependent trade-off between improved local detection and governance complexity. By separating territorial fairness auditing from performance-based model comparison, this study provides an evidence-based framework for reasoning about regionalization in large-scale educational risk prediction.

Authors: Francisco, TV; Barros, AN; Roque, FV; Paulino, T; Schmidt, A

• Publication status: Published
• Peer review status: Peer reviewed
• Publisher: Springer
• Host title: Artificial Intelligence in Education. AIED 2026
• Pages: 225-239
• Series: Lecture Notes in Computer Science
• Series number: 16586
• Publication date: 2026-06-27
• Acceptance date: 2026-06-27
• Event title: Artificial Intelligence in Education (AIED 2026)
• Event location: Seoul, South Korea
• Event website: https://www.aied-conference.org/2026
• Event start date: 2026-06-27
• Event end date: 2026-07-03
• DOI: 10.1007/978-3-032-29773-0_16
• EISSN: 1611-3349
• ISSN: 0302-9743
• EISBN: 9783032297730
• ISBN: 9783032297723
• Language: English
• Keywords: learning analytics; academic risk prediction; algorithmic fairness; territorial equity