NMTAFE Early Dropout-Risk Detection (AT4)
Predicts whether a student is at risk of withdrawing or failing, using only the first 90 days of a course.
Model
- Final model: XGBoost (56 features, 218 trees, best = 188, depth 5)
- LSTM (4 layers, 64 units, daily clicks over 91 days) tested in a hybrid; weight search chose 100% XGBoost
- Decision threshold: 0.38
Data
Open University Learning Analytics Dataset (OULAD), Kuzilek et al. (2017), CC-BY 4.0.
Test results (unseen students)
| Metric | Result | Target |
|---|---|---|
| Accuracy | 0.753 | > 0.90 |
| Precision | 0.651 | > 0.90 |
| Recall | 0.828 | > 0.90 |
| ROC-AUC | 0.858 | - |
Files
- xgb_dropout.json – XGBoost model
- lstm_dropout.pt – LSTM weights (PyTorch)
- scaler.pkl – StandardScaler (fitted on training data)
- config.json – feature list, threshold, hybrid weight, settings
Intended use and limitations
Weekly risk list for student support staff. Every flag must be reviewed by a person. Trained on UK data (2013–14), so it must be retrained on local data before real use. Sensitive features (gender, disability, age, IMD, region) are not model inputs.
- Downloads last month
- 28