Machine learning-based predictive model for optimal A-level combination selection using O-level academic performance in Rwandan high schools

The selection of A-level subject combinations in Rwandan high schools significantly influencesstudents' academic and career trajectories. However, the existing manual selection process,influenced by subjective factors such as parental, peer pressure, and institutional preferences, oftenresults in misalignments between students' abilities and chosen combinations. This study proposesthe first machine learning-based predictive model tailored for Rwanda's education system tooptimize A-level combination selection using students' O-level academic performance data. Adataset comprising 2,614 students from selected schools was examined using various classificationalgorithms, including Support Vector Classifier (SVC), Random Forest (RF), and Decision Tree (DT)models. To handle class imbalance, the Synthetic Minority Over-Sampling Technique (SMOTE) wasemployed. The models' performance was measured based on accuracy, precision, recall, andF1-score, with SVC achieving the highest performance, attaining 97.51% accuracy, 97.43% precision,97.44% F1-score, and 97.49% recall. A k-fold cross-validation approach (k=10) was applied to validatemodel robustness (SVC Mean Accuracy (10-Fold CV): 97.61%, Random Forest Mean Accuracy (10-FoldCV): 84.94%, and Decision Tree Mean Accuracy (10-Fold CV): 67.15%). Statistical significance tests(ANOVA) confirmed the reliability of SVC’s superior performance (ANOVA F-statistic: 260.8576, ANOVAp-value: 0.000000). SVC performance is significantly better than RF and DT (p < 0.05). The studydemonstrates the potential of predictive analytics in enhancing student placement by providingobjective, data-driven recommendations. Future improvements include incorporating deep learningmodels, integrating student feedback, and conducting longitudinal studies to assess the impact ofmachine- assisted selection over time.