Student Performance Prediction (Machine Learning)

This project develops regression models to predict final student grades (G3) using demographic, behavioral, and academic features from a public dataset.

The primary objective is to evaluate predictive performance under two real-world intervention scenarios:

1. With prior-term grades (G1, G2)
   High predictive accuracy, usable later in the academic year.

2. Without prior-term grades
   Lower accuracy, but useful for early risk detection and intervention planning.

---

Dataset

• Source: UCI Machine Learning Repository
• 395 student records
• 35 features
• Target variable: Final grade (G3)
• 80/20 train-test split

---

Modeling Approach

• Feature engineering with custom transformer
• ColumnTransformer pipelines for:
  • Numeric features
  • Categorical (One-Hot Encoding)
  • Ordinal encoding
• Cross-validation (3-fold)
• Grid search hyperparameter tuning (SVR)

---

Models Evaluated

• Linear Regression
• Lasso Regression
• Support Vector Regression (SVR)

---

Test Set Performance

With Prior-Term Grades

• RMSE: 2.142
• R²: 0.776

Without Prior-Term Grades

• RMSE: 4.154
• R²: 0.158

---

Key Insights

• Prior-term grades dominate predictive power.
• Early-year prediction is substantially more difficult.
• Feature engineering and structured pipelines significantly impacted model performance.
• Even lower-accuracy early models can be operationally valuable for identifying at-risk students.

---

Repository Structure

notebooks/student_performance_ml_portfolio.ipynb

reports/student_performance_executive_summary.pdf

Business Implications

• Use grade-inclusive model for mid-year intervention targeting.
• Use early-warning model to flag high-risk students before term grades are available.
• Focus interventions on attendance and study-time support.