Final Maths grade, estimated.
Describe a student and the selected model estimates their final grade on the 0–20 scale. Below 10 is marked for a supportive review by staff, never for an automatic decision.
A whole class at once.
Upload a CSV with one student per row, using the same column names as the dataset. Rows with both G1 and G2 use the later-stage model; the others use the main model.
No file yet. Try the template with a few fictional students.
| Row | Model | Predicted G3 | Status | Notes |
|---|
Two models, one rule.
Notebook 03 compared Linear Regression, Ridge, Decision Tree, Random Forest and Gradient Boosting on the same 316 training students with five-fold cross-validation, and kept the lowest mean absolute error in each scenario. Random Forest won both. The app switches between them only by what you know about the student.
Test figures are recomputed from the saved models on the same 79 held-out students each time the app starts, and match Notebook 04. That test set has already been inspected, so it is not used for further tuning.
Keep in mind
- The main model misses 17 of the 27 test students who scored below 10. An unflagged student can still need help.
- At the later stage, reusing G2 as the prediction has a lower error (1.000) than the forest.
- The data come from two Portuguese schools, so results may not transfer elsewhere.
- The causes of the 38 recorded zero grades are unknown, and the meaning of the top failures code is unconfirmed.
- Predictions describe associations. They do not show what an intervention would change.