LassoNet

LassoNet provides specialized models for different tasks:

• Regression: LassoNetRegressor
• Classification: LassoNetClassifier
• Cox regression: LassoNetCoxRegressor
• Interval-censored Cox regression: LassoNetIntervalRegressor

For cross-validation, use the versions ending in CV (e.g., LassoNetRegressorCV).

Install the lassonet package using pip to access the LassoNet regression, classification, and Cox regression models.

pip install lassonet

LassoNet follows scikit-learn standards. You can use LassoNetClassifierCV or LassoNetRegressorCV for cross-validated training. Always normalize your data before passing it to the model, as it uses neural networks under the hood.

from lassonet import LassoNetClassifierCV 
# or LassoNetRegressorCV

model = LassoNetClassifierCV() 
path = model.fit(X_train, y_train)
print("Best model scored", model.score(X_test, y_test))
print("Lambda =", model.best_lambda_)

After finding selected features via the regularization path, you can train a dense model (without the L1 skip connection) specifically on those features by setting dense_only=True in the .fit() method.

# Assuming 'selected' is a boolean mask or index array from a path checkpoint
model.fit(X_train[:, selected], y_train, dense_only=True)
print(model.score(X_test[:, selected], y_test))

Models ending in CV (e.g., LassoNetRegressorCV) perform 5-fold cross-validation by default. For each fold, a path is trained, and the best regularization value is chosen to maximize the average score. The model is then retrained on the entire training set.

Key attributes for the best model:

• best_lambda_: The optimal regularization parameter.
• best_selected_: The optimal subset of features.
• best_cv_score_: The best cross-validation score.

You can also use model.plot_cv() to inspect the cross-validation results.

model = LassoNetRegressorCV()
model.fit(X_train, y_train)
model.score(X_test, y_test)
# Use model.plot_cv() for more info

The base models allow you to train along a regularization path using the .path() method. This returns a list of checkpoints. You can visualize the path with plot_path() and then load a specific model from the path using .load() with a state dictionary.

from lassonet import LassoNetRegressor, plot_path

model = LassoNetRegressor()
# Returns a list of checkpoints
path = model.path(X_train, y_train, return_state_dicts=True)

# Visualize the path
plot_path(model, X_test, y_test)

# Choose best_id based on the plot and load it
model.load(path[best_id].state_dict)
print(model.score(X_test, y_test))

Stability selection is the recommended way to identify the most important features by running the model multiple times on different random subsamples. Use model.stability_selection() to perform this.

model = LassoNetRegressor()
# Returns several diagnostic objects
oracle, order, wrong, paths, prob = model.stability_selection(X_train, y_train)

# 'order' sorts features by decreasing importance
# 'prob' is the probability a feature is selected at each lambda value

LassoNet includes utility functions for visualizing the learned paths and model structures:

• lassonet.plot_path: Used to plot the regularization paths.
• lassonet.lassonet_path: Used to visualize the LassoNet path.

LassoNet provides several model classes for different machine learning tasks, including cross-validated versions for automated hyperparameter tuning:

Regression

• lassonet.LassoNetRegressor: Standard LassoNet regressor.
• lassonet.LassoNetRegressorCV: LassoNet regressor with built-in cross-validation.

Classification

• lassonet.LassoNetClassifier: Standard LassoNet classifier.
• lassonet.LassoNetClassifierCV: LassoNet classifier with built-in cross-validation.

Survival Analysis (Cox)

• lassonet.LassoNetCoxRegressor: LassoNet regressor for Cox proportional hazards models.
• lassonet.LassoNetCoxRegressorCV: LassoNet Cox regressor with built-in cross-validation.

When initializing LassoNet models, the following parameters are available:

• hidden_dims: Number of neurons in each hidden layer (e.g., (100,) or (10, 10)).
• path_multiplier: Number of lambda values to compute on the path. Lower values increase precision but take more time (e.g., 1.01).
• lambda_start: Starting value for regularization. Defaults to `