An end-to-end Machine Learning project that predicts a student's final academic grade (G3) using a Random Forest Regressor, deployed as an interactive Streamlit web app.
Schools and tutors often want early signals of which students may need extra support. This project explores whether a student's final grade can be predicted from study habits, attendance, and prior performance.
UCI Student Performance Dataset (student-mat.csv) — Portuguese secondary
school students, including demographic, academic, and lifestyle attributes.
Two models were trained and compared:
| Model | MAE | R² |
|---|---|---|
| Linear Regression | 1.34 | 0.78 |
| Random Forest Regressor | 1.05 | 0.87 |
Random Forest was selected as the final model after hyperparameter tuning
(n_estimators=100, max_depth=5, selected via 5-fold cross-validation).
Features used: Study Time, Past Failures, Absences, G1 (first period grade), G2 (second period grade).
Final Model Performance (test set):
- R² Score: 0.8592
- MAE: 1.0952
Most predictions track closely with actual grades, particularly for mid-to-high performers (8–16 range). The model is less reliable at the extremes — students with very low actual grades (near 0) are sometimes over-predicted.
A common criticism of this dataset: G1 and G2 are prior-period grades, not lifestyle factors. To test how much they drive the result, a second model was trained using only Study Time, Failures, and Absences:
| Model Version | R² | MAE |
|---|---|---|
| With G1, G2 | 0.8592 | 1.0952 |
| Without G1, G2 | 0.1032 | 3.4223 |
This is the most important finding of this project: G1 and G2 account for the vast majority of this model's predictive power. Without them, study time/failures/absences alone explain very little of the variation in final grades.
In practice, this model is better understood as a grade trend extrapolator (estimating G3 from prior academic trajectory) rather than a model that predicts outcomes purely from lifestyle habits.
Python · Pandas · NumPy · Scikit-Learn · Streamlit · Joblib
StudentPerformancePredictor_Final/
├── Dataset/
├── Model/
├── src/
│ ├── data_preprocessing.py
│ ├── train_model.py
│ ├── evaluate.py
│ └── main.py
├── StudentPerformancePredictor_Final.ipynb
├── app.py
├── requirements.txt
└── README.md
git clone https://github.com/mukundan1012-creator/StudentPerformancePredictor_Final.git
cd StudentPerformancePredictor_Final
pip install -r requirements.txt
streamlit run app.pyTo retrain the models from scratch:
cd src
python main.pyI would test additional engineered features (e.g., parental education, weekly alcohol consumption) on the no-grades model to see if the R² gap can be narrowed, and add a feature-importance plot to visualize exactly how much weight G1/G2 carry relative to other inputs.
Mukundan D B.E. Electronics and Communication Engineering