机器学习课程设计
An end-to-end machine-learning project that analyses student attributes and predicts academic progression outcomes.
26p
course report
~1.3
labelled MAE
06
base model families
01 / Overview
The machine-learning coursework predicts a student’s final grade from demographic, family, study-habit and prior-term variables while controlling for leakage. The strongest saved model is labelled with an MAE of approximately 1.3.
Objectives
ML workflow
The page focuses on leakage control, feature engineering and model comparison.
Academic, demographic, family and lifestyle attributes.
Prior-grade means, differences, alcohol total and parental education.
Ridge, Random Forest, ExtraTrees, XGBoost, LightGBM and CatBoost.
Ridge meta-learner combining base regressors.
MAE, RMSE, R² and saved best model.
02 / Package highlights
03 / Model design
The project creates interaction and context features such as prior-grade averages, grade change, combined alcohol consumption and parental education. The original notebook explicitly notes the importance of removing G1 and G2 from the feature set in some experiments to avoid target leakage into the final-grade prediction.
The target is tabular and moderately sized. A stack of linear, bagged-tree, extremely randomised tree and gradient-boosting models captures different error patterns. A Ridge meta-learner keeps the combination stable and passes the original features through to the second stage.
The workflow compares MAE, RMSE and R² across models and saves the best-performing estimator. The resulting artifact is labelled “Ultimate_Student_Model_MAE_1.3”, indicating the project’s approximate mean absolute error target.
04 / Key code
Only verified source excerpts and pseudocode taken from the project papers are shown.
Original Python excerpt for feature construction and leakage control.
student['G1_G2_mean'] = (student['G1'] + student['G2']) / 2
student['G1_G2_diff'] = student['G2'] - student['G1']
student['total_alcohol'] = student['Dalc'] + student['Walc']
student['parents_edu'] = student['Medu'] + student['Fedu']
student['want_higher'] = (student['higher'] == 'yes').astype(int)
student = pd.get_dummies(student, drop_first=True)
labels = student['G3']
# Remove the raw prior grades when testing leakage-safe models
features = student.drop(['G3', 'G1', 'G2'], axis=1)
X_train, X_test, y_train, y_test = train_test_split(
features, labels, test_size=0.25, random_state=42
)Original Python excerpt showing the model stack and Ridge meta-learner.
base_models = [
('lr', LinearRegression()),
('rf', RandomForestRegressor(n_estimators=300, random_state=42)),
('et', ExtraTreesRegressor(n_estimators=300, random_state=42)),
('xgb', XGBRegressor(
n_estimators=800, learning_rate=0.02,
max_depth=5, random_state=42
)),
('lgb', LGBMRegressor(
n_estimators=800, learning_rate=0.02,
num_leaves=31, random_state=42, verbose=-1
)),
('cat', CatBoostRegressor(
iterations=800, learning_rate=0.02,
depth=6, random_seed=42, verbose=False
)),
]
final_model = StackingRegressor(
estimators=base_models,
final_estimator=Ridge(alpha=1.0),
passthrough=True,
cv=10,
n_jobs=-1,
)Evaluation notes
Limitations / next steps
05 / Downloads
Open the original report, presentation or source archive. PDF cover pages with private details were removed.
Sanitized 26-page report without the student cover page.