<![CDATA[The delayed identification of students at risk of academic underperformance frequently undermines the effectiveness of pedagogical interventions. This limitation arises because most traditional models for predicting student performance depend on exhaustive end-of-semester datasets, engendering a latency issue wherein insights emerge too late for timely remediation. To overcome this, we propose an Explainable Early Warning System that forecasts students' Final Year Mathematics Assessment scores using exclusively mid-semester data: Daily Assessments and Mid-Semester Assessments. By utilizing an augmented dataset of 68 students, hybrid data augmentation to fix class imbalance, and a Restricted Random Forest model to prevent overfitting, our method achieves a strong 92.3% classification accuracy on unseen test data. Remarkably, it achieves 100% Recall for the 'Need Guidance' class, ensuring no at-risk students are overlooked. Furthermore, SHAP analysis reveals that, beyond midterm scores, consistency in specific daily tasks, particularly Daily Assessment Chapter 4 significantly impacts failure risk. In conclusion, combining data augmentation with explainable machine learning transforms predictions into actionable pedagogical insights, empowering teachers to execute precise interventions three months prior to the final exam.]]>
Copyrights © 2026
