Data-Driven Prediction of Student Failure Using Ensemble Machine Learning Models
TL;DR
This study employs a hybrid ensemble stacking approach to improve predictions of student failure based on AKMI scores and underscores the potential of ensemble learning models in leveraging AKMI data for early identification of students at academic risk.
Abstract
Education is a fundamental factor in shaping students' future success; however, accurately predicting student failure remains challenging due to many influencing variables. The Indonesian Madrasah Competency Assessment (AKMI), developed by the Ministry of Religious Affairs (Kemenag RI), evaluates student competencies across four literacy domains: reading, numeracy, scientific, and sociocultural literacy. While AKMI offers a comprehensive assessment framework, the integration of advanced analytical methods such as Machine Learning (ML) can significantly enhance the predictive accuracy of student performance and support data-driven educational decisions. This study employs a hybrid ensemble stacking approach to improve predictions of student failure based on AKMI scores. Two stacking models were constructed: Stacking Classifier (1), which integrates K-Nearest Neighbors (KNN) and Decision Tree (DT) with Naïve Bayes as the meta-learner; and Stacking Classifier (2), which combines Random Forest (RF) and Decision Tree (DT) with Naïve Bayes. These models were evaluated using key performance indicators, including accuracy, precision, recall, F1-score, and computational time. The results show that Stacking Classifier (1) achieved the highest accuracy at 95%, whereas Stacking Classifier (2) demonstrated a favorable trade-off between accuracy (90%) and computational efficiency (1.306 seconds, compared to 734.951 seconds for Classifier 1). Both classifiers were tested with integrated demographic and academic datasets, revealing that incorporating both types of data enhances predictive performance. These findings underscore the potential of ensemble learning models in leveraging AKMI data for early identification of students at academic risk. Furthermore, the application of ML in madrasah education offers valuable insights for educators and policymakers, facilitating the development of targeted interventions, optimized learning strategies, and improved educational outcomes.