ML Python Lab 2 — California Housing Regression

This lab demonstrates a compact, reproducible workflow for tabular regression with scikit-learn using the California Housing dataset. It provides:

• Reusable pipelines with preprocessing and models
• Hyperparameter search via GridSearchCV + KFold
• Evaluation with RMSE, MAE, and R²
• A Jupyter notebook to explore experiments and figures

Project Structure

ml_python/lab2/
├─ src/
│  └─ housing.py            # Pipelines, training, evaluation utilities
├─ notebooks/
│  ├─ experiment.ipynb      # Main notebook to run experiments
│  └─ figures/              # Generated figures (kept in repo)
├─ requirements.txt         # Python dependencies
├─ .gitignore               # Ignore caches, data, models, etc.
└─ README.md                # You are here

Requirements

• Python 3.9+ (3.10/3.11 recommended)
• See pinned packages in requirements.txt:
  • scikit-learn, numpy, pandas, matplotlib, jupyterlab
  • dev tools: ruff, nbstripout

Optional but recommended:

• A virtual environment (e.g., venv, conda)

Quickstart

1. Create and activate a virtual environment

python -m venv .venv
source .venv/bin/activate   # Windows: .venv\Scripts\activate

2. Install dependencies

pip install -r requirements.txt

3. Launch JupyterLab and open the notebook

jupyter lab

Open notebooks/experiment.ipynb and run the cells.

Using the Library (Python API)

You can also run experiments from Python using the utilities in src/housing.py.

Available models (keys in MODELS):

• simple_elastic: ElasticNet with standard scaling
• poly_elastic: PolynomialFeatures (degree=2) + ElasticNet
• knn: KNeighborsRegressor with standard scaling

Each model includes a param_grid for GridSearchCV.

Example workflow:

from src.housing import set_seed, load_dataset, train, eval

# 1) Reproducibility
set_seed(1)

# 2) Load data
X_train, X_test, y_train, y_test = load_dataset(test_size=0.2, random_state=1)

# 3) Train with cross-validated grid search
search = train(
    model="simple_elastic",    # or "poly_elastic", "knn"
    X_train=X_train,
    y_train=y_train,
    cv_splits=5,
    random_state=1,
)

# 4) Evaluate on the test set
results = eval(search, X_test, y_test)
print("Best estimator:", results["estimator"])  # the fitted model inside the pipeline
print("RMSE:", results["rmse"])                 # note: scikit-learn returns MSE by default; RMSE ~ sqrt(MSE)
print("MAE:", results["mae"])                  
print("R^2:", results["r2"])                   

Notes:

• The returned search is a fitted GridSearchCV. The best full pipeline is at search.best_estimator_.
• eval returns predictions and metrics for convenience.

Reproducibility & Configuration

• Randomness: Use set_seed(seed) and pass random_state to load_dataset and train.
• Cross-validation: Configure with cv_splits in train (default 5).
• Hyperparameters: Adjust param_grid per model inside src/housing.py or pass a custom configuration by extending the code.

Notebook Hygiene

• nbstripout is included to help keep notebooks lightweight. To activate in this repo:

nbstripout --install

This installs a Git filter that strips large outputs on commit. You can reverse with nbstripout --uninstall.

Data & Artifacts

• The dataset is fetched programmatically via sklearn.datasets.fetch_california_housing, so no local data is required.
• .gitignore excludes typical data/model artifacts (data/, models/, *.pkl, *.csv, etc.) to keep the repo clean.

Troubleshooting

• If Jupyter cannot find the kernel, ensure your virtual environment is active when installing and launching JupyterLab.
• Version conflicts: Reinstall with a fresh environment and pip install -r requirements.txt.
• Long grid searches: Reduce cv_splits, narrow param_grid, or try a smaller model first.

License

No license specified. If you plan to share or publish, consider adding a LICENSE file.