Preface

Kaggle ML Learning is where I ran the full competition loop for the first time: load Iowa house data, pick a lean feature set, measure with a holdout, refit on all labels, predict the test file, ship submission.csv.

It started as the Kaggle Learn Machine Learning Competitions exercise and lives on as a personal archive of that first end-to-end submit.

kaggle-machine-learning-learning on GitHub

https://github.com/lucaimbalzano/kaggle-machine-learning-learning

The problem

Predict SalePrice for homes in the Home Data for ML Course competition. Classic tabular regression: many columns, a public leaderboard, and an easy trap — treating training error as if it were validation.

The goal for this exercise was not a fancy stack. It was discipline: same feature schema on train and test, a real holdout score before submit, then a full-data refit for the CSV Kaggle expects.

Results

Metric	Value	Meaning
Holdout MAE	~21,857	Fit on a train split, scored on unseen validation rows
Train MAE (full fit)	~8,300	Scored on data it just saw — optimistic, not a validation claim
Output	submission.csv	Id + SalePrice for the competition test set

The gap between those MAEs is the lesson: a good programmer trusts the holdout, not the flattering number.

House Prices — first Kaggle competition loop with Random Forest and holdout MAE

Folder structure

Intentionally small — one exercise folder, one notebook, room to grow.

exercise-machine-learning-competitions

exercise-machine-learning-competitions.ipynb

README.md

LICENSE

Under the hood

Two snippets that matter more than the algorithm choice: measure before you submit, and keep the feature contract identical on test.

Holdout first — then refit for the leaderboard

exercise-machine-learning-competitions.ipynb

features = [
    'LotArea', 'YearBuilt', '1stFlrSF', '2ndFlrSF',
    'FullBath', 'BedroomAbvGr', 'TotRmsAbvGrd',
]
X = home_data[features]
y = home_data.SalePrice

train_X, val_X, train_y, val_y = train_test_split(X, y, random_state=1)

rf_model = RandomForestRegressor(random_state=1)
rf_model.fit(train_X, train_y)
rf_val_mae = mean_absolute_error(rf_model.predict(val_X), val_y)
# holdout MAE ≈ 21,857 — the number you can defend

rf_model_on_full_data = RandomForestRegressor(random_state=1)
rf_model_on_full_data.fit(X, y)  # all labels only after you have a local score

Same columns on test — competition CSV shape

exercise-machine-learning-competitions.ipynb

test_X = test_data[features]  # identical schema to train X
test_preds = rf_model_on_full_data.predict(test_X)

output = pd.DataFrame({'Id': test_data.Id, 'SalePrice': test_preds})
output.to_csv('submission.csv', index=False)

Small details, strong habit: one features list, random_state for reproducibility, and a submission frame that matches the competition contract exactly.

Tech stack

Open source & links

View on GitHub

https://github.com/lucaimbalzano/kaggle-machine-learning-learning

Kaggle notebook

https://www.kaggle.com/code/francescoimbalzano/exercise-machine-learning-competitions?scriptVersionId=241262910

Next stops on this track: missing values, categoricals, and a stronger booster baseline — still with the same holdout-before-submit muscle memory.

Kaggle ML Learning

First Kaggle competition loop — House Prices with a Random Forest, honest holdout MAE, and a clean submission.csv.