Low-Connectivity Learning Analytics: Lightweight Predictive Models for School Dropout Prevention in Mozambique.

Abstract School dropout remains one of the greatest challenges to human development in Sub-Saharan Africa, particularly in Mozambique. Traditional Learning Analytics approaches require high connectivity and real-time interaction, which are infeasible in rural and low-resource environments. This article presents the design, validation, and evaluation of a Machine Learning (ML) model optimized for predicting dropout risk using exclusively asynchronous, low-volume data. By leveraging only three input features—frequency of task submission (FET), periodic test performance (DAP), and absenteeism tendency (TF)—we achieved high predictive accuracy with minimal computational cost. The Random Forest Classifier achieved an AUC of 0.95 and a critical Recall (Sensitivity) of 92.1% ± 1.8%, significantly outperforming the Logistic Regression baseline (AUC 0.88). The model’s runtime was under 150 milliseconds, and its memory consumption was below 50MB, demonstrating its viability for offline-first and SMS-based architectures. Partial Dependence Analysis ensured interpretability, revealing Performance (DAP) and Task Submission (FET) as the strongest predictors. This study contributes to Educational Data Mining (EDM) by demonstrating a lightweight, equitable AI solution that is viable under extreme resource constraints, directly addressing digital inequities and providing actionable recommendations for the Mozambican Ministry of Education (MINEDH).