<![CDATA[Abstract: This research aims to develop an early detection system for at-risk students using machine learning based on data from the Learning Management System (LMS) rumahilmu.org. The system was designed for the Information Systems Study Programs at the University of Bengkulu, analyzing data from 459 student enrollments across five courses. A total of 37–76 features were extracted from LMS activities to predict students likely to score below the 30th percentile at three strategic time points (25%, 50%, and 75% of the semester). This study implemented a per-class optimization approach, testing 11 algorithms to find the best model for each course. The results showed that no single algorithm was universally superior; the most effective models varied for each course, with Gaussian Process, Logistic Regression, and Voting Classifier being the most frequently chosen. However, evaluation on the test data revealed significant challenges: despite high cross-validation scores (F1-score > 0.80), overfitting and performance degradation occurred. The most critical finding was the model's low capability in detecting the 'At-Risk' minority class, with the Recall (At-Risk) metric reaching 0.00 in 8 out of 15 scenarios. The best detection performance was achieved in the Statistics & Probability course with a Recall of 0.50. The implemented system, featuring a 3-tier architecture (FastAPI and React), provides an interactive dashboard, but its predictive effectiveness for early detection is limited by small and imbalanced datasets.]]>
Copyrights © 2025
