Work

An energy-efficient and explainable green AI framework for student dropout prediction with a rule-based decision support module

Year: 2026
Type: article
Abstract: Student dropout is a persistent issue in massive open online and blended learning environments, and identifying vulnerable learners early is essential. Although deep learning (DL) techniques have increased prediction accuracy, they often have substantial computational overheads and insufficiently transparent operations, thus hindering their implementation in resource-scarce educational settings. This work presents a green AI and explainable learning approach for the prediction of student dropout using the Open University Learning Analytics Dataset (OULAD). A small multilayer perceptron model, GreenMLP, is constructed to capture the behavioral pattern of learners by analyzing their demographic information and virtual learning environment logs. The GreenMLP model is trained under a low-resource setting on CPUs only, where energy expenditure during training is tracked using CodeCarbon, allowing the direct calculation of the carbon footprint. In addition, eXplainable AI (XAI) approaches such as SHAP and LIME are used to provide global and local explanations for the prediction task. The experimental results on the OULAD demonstrate that the proposed GreenMLP model achieves an AUC of 0.8361 and an accuracy of 0.7584 while using only 2,753 parameters and producing 2.08 × 10⁻⁵ kg of CO₂ during training. The explainability analysis reveals that student engagement features, particularly total and peak interaction levels, dominate dropout risk, whereas demographic factors play secondary roles. The rule-based decision layer reduces false alerts and increases intervention precision. These results show that energy-efficient and XAI methods can achieve competitive results and help maintain transparency and sustainability in the assessment of dropout risk in digital education.
Language: English
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7226007462
FWCI:
Cites: 2
Cited by:
Related to: 10