In progressSeptember 2026 – PresentActive, with weekly supervisor meetings
Checkpoint Early-Warning Prediction of Student Performance
Advanced Technology and Artificial Intelligence Center, AURAK
Supervisor: Prof. Khouloud Salameh
A system that estimates each student's risk of failing or withdrawing, final letter grade and term GPA at weeks 0, 4, 8, 12 and 16 of a term, so advisors can step in early. It uses only the information available at each checkpoint, and runs alongside the team's systematic literature review of student performance prediction. Because AURAK records need ethics approval, it has so far been evaluated on three public datasets.
What I work on
- Designed and built the full prototype, using only information available at each checkpoint: final exams are excluded and coursework scores count only once released
- Compared explicit baselines, interpretable models (elastic net, proportional-odds and L2 logistic regression) and LightGBM under term-forward splits, keeping a model only if it showed skill on a separate calibration term
- Calibrated probabilities (Platt, isotonic and temperature scaling) and added conformalized prediction intervals, advisor-capacity risk tiers and a subgroup fairness audit (recall, calibration gap, ABROCA)
- Evaluated on OULAD, KU Leuven ALPACAS and SED. On OULAD, PR-AUC rose from 0.58 at week 0 to 0.77 at week 16, and the evaluation documents interval under-coverage and calibration loss when the test term differed from the training terms
- Built a local FastAPI prediction service and an accessible advisor dashboard, and contributed the model comparison and study review tables to the literature review
Pythonscikit-learnLightGBMConformal PredictionFastAPIModel Evaluation