asgl

Introduction

The asgl package is a versatile and robust tool designed for fitting a variety of regression models, including linear regression, quantile regression, logistic regression and various penalized regression models such as Lasso, Ridge, Group Lasso, Sparse Group Lasso, and their adaptive variants. The package is especially useful for simultaneous variable selection and prediction in both low and high-dimensional frameworks.

The primary class available to users is the Regressor class, which is detailed later in this document.

asgl is based on cutting-edge research and methodologies, as outlined in the following papers:

• Adaptive Sparse Group Lasso in Quantile Regression
• asgl: A Python Package for Penalized Linear and Quantile Regression

For a practical introduction to the package, users can refer to the user guide notebook available in the GitHub repository. Additional accessible explanations can be found on Towards Data Science: Sparse Group Lasso, Towards Data Science: Adaptive Lasso and Towards Data Science: Quantile regression.

Dependencies

asgl requires:

• Python >= 3.9
• cvxpy >= 1.2.0
• numpy >= 1.20.0
• scikit-learn >= 1.0
• pytest >= 7.1.2

User installation

The easiest way to install asgl is using pip:

pip install asgl

Testing

After installation, you can launch the test suite from the source directory (you will need to have pytest >= 7.1.2 installed) by runnig:

pytest

What’s new?

2.1.1

Now the intercept term appears in the intercept_ attribute instead of being part of the coef_ attribute.

2.1.0

The latest release of the asgl package, version 2.1.0, introduces powerful enhancements for logistic regression models. Users can now easily tackle binary classification problems by setting model='logit'. For more granular control, specify model='logit_raw' to retrieve outputs before logistic transformation, or model='logit_proba' for probability outputs. Additionally, this update includes the implementation of ridge and adaptive ridge penalizations, accessible via penalization='ridge' or 'aridge', allowing for more flexible model tuning.

2.0.0

With the release of version 2.0, the asgl package has undergone significant enhancements and improvements. The most notable change is the introduction of the Regressor object, which brings full compatibility with scikit-learn. This means that the Regressor object can now be used just like any other scikit-learn estimator, enabling seamless integration with scikit-learn’s extensive suite of tools for model evaluation, hyperparameter optimization, and performance metrics.

Key updates include:

• Scikit-learn Compatibility: The Regressor class is now fully compatible with scikit-learn. Users can leverage functionalities such as sklearn.model_selection.GridSearchCV for hyperparameter tuning and utilize various scikit-learn metrics and utilities to assess model performance.

• Deprecation of ASGL class: The old ASGL class is still included in the package for backward compatibility but is now deprecated. It will raise a DeprecationWarning when used, as it is no longer supported and will be removed in future versions. Users are strongly encouraged to transition to the new Regressor class to take advantage of the latest features and improvements.

For users currently utilizing the ASGL class, we recommend switching to the Regressor class to ensure continued support and access to the latest functionalities.

Key features:

The Regressor class includes the following list of parameters:

• model: str, default=‘lm’
  • Type of model to fit. Options are ‘lm’ (linear regression), ‘qr’ (quantile regression), ‘logit’ (logistic regression for binary classification, output binary classification), ‘logit_proba’ (logistic regression for binary classification, output probability) and ‘logit_raw’ (logistic regression for binary classification, output score before logistic).
• penalization: str or None, default=‘lasso’
  • Type of penalization to use. Options are ‘lasso’, ‘ridge’, ‘gl’ (group lasso), ‘sgl’ (sparse group lasso), ‘alasso’ (adaptive lasso), ‘aridge’, ‘agl’ (adaptive group lasso), ‘asgl’ (adaptive sparse group lasso), or None.
• quantile: float, default=0.5
  • Quantile level for quantile regression models. Valid values are between 0 and 1.
• fit_intercept: bool, default=True
  • Whether to fit an intercept in the model.
• lambda1: float, default=0.1
  • Constant that multiplies the penalization, controlling the strength. Must be a non-negative float i.e. in [0, inf). Larger values will result in larger penalizations.
• alpha: float, default=0.5
  • Constant that performs tradeoff between individual and group penalizations in sgl and asgl penalizations. alpha=1 enforces a lasso penalization while alpha=0 enforces a group lasso penalization.
• solver: str, default=‘default’
  • Solver to be used by cvxpy. Default uses optimal alternative depending on the problem. Users can check available solvers via the command cvxpy.installed_solvers().
• weight_technique: str, default=‘pca_pct’
  • Technique used to fit adaptive weights. Options include ‘pca_1’, ‘pca_pct’, ‘pls_1’, ‘pls_pct’, ‘lasso’, ‘ridge’, ‘unpenalized’, and ‘sparse_pca’. For low dimensional problems (where the number of variables is smaller than the number of observations) the usage of the ‘unpenalized’ or ‘ridge’ weight_techniques is encouraged. For high dimensional problems (where the number of variables is larger than the number of observations) the default ‘pca_pct’ is encouraged.
• individual_power_weight: float, default=1
  • Power to which individual weights are raised. This parameter only has effect in adaptive penalizations. (‘alasso’ and ‘asgl’).
• group_power_weight: float, default=1
  • Power to which group weights are raised. This parameter only has effect in adaptive penalizations with a grouped structure (‘agl’ and ‘asgl’).
• variability_pct: float, default=0.9
  • Percentage of variability explained by PCA, PLS, and sparse PCA components. This parameter only has effect in adaptiv penalizations where weight_technique is equal to ‘pca_pct’, ‘pls_pct’ or ‘sparse_pca’.
• lambda1_weights: float, default=0.1
  • The value of the parameter lambda1 used to solve the lasso model if weight_technique='lasso'
• spca_alpha: float, default=1e-5
  • Sparse PCA parameter. This parameter only has effect if weight_technique='sparse_pca'See scikit-learn implementation for more details.
• spca_ridge_alpha: float, default=1e-2
  • Sparse PCA parameter. This parameter only has effect if weight_technique='sparse_pca'See scikit-learn implementation for more details.
• individual_weights: array or None, default=None
  • Custom individual weights for adaptive penalizations. If this parameter is informed, it overrides the weight estimation process defined by parameter weight_technique and allows the user to provide custom weights. It must be either None or be an array with non-negative float values and length equal to the number of variables.
• group_weights: array or None, default=None
  • Custom group weights for adaptive penalizations. If this parameter is informed, it overrides the weight estimation process defined by parameter weight_technique and allows the user to provide custom weights. It must be either None or be an array with non-negative float values and length equal to the number of groups (as defined by group_index)
• tol: float, default=1e-4
  • Tolerance for coefficients to be considered zero.
• weight_tol: float, default=1e-4
  • Tolerance value used to avoid ZeroDivision errors when computing the weights.

Examples

Example 1: Linear Regression with Lasso.

from sklearn.datasets import make_regression
from sklearn.model_selection import train_test_split
from sklearn.metrics import mean_squared_error
from asgl import Regressor

X, y = make_regression(n_samples=1000, n_features=50, n_informative=25, bias=10, noise=5, random_state=42)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=250)

model = Regressor(model='lm', penalization='lasso', lambda1=0.1)
model.fit(X_train, y_train)

predictions = model.predict(X_test)
mse = mean_squared_error(predictions, y_test)

This example illustrates how to:

• Generate synthetic regression data.
• Split the data into training and testing sets.
• Create a Regressor object configured for linear regression with Lasso penalization.
• Fit the model to the training data.
• Make predictions on the test data.
• Evaluate the model’s performance using mean squared error.

Example 2: Quantile regression with Adaptive Sparse Group Lasso.

Group-based penalizations like Group Lasso, Sparse Group Lasso, and their adaptive variants, assume that there is a group structure within the regressors. This structure can be useful in various applications, such as when using dummy variables where all the dummies of the same variable belong to the same group, or in genetic data analysis where genes are grouped into genetic pathways.

For scenarios where the regressors have a known grouped structure, this information can be passed to the Regressor class during model fitting using the group_index parameter. This parameter is an array where each element indicates the group at which the associated variable belongs. The following example demonstrates this with a synthetic group_index. The model will be optimized using scikit-learn’s RandomizedSearchCV function.

import numpy as np
from sklearn.datasets import make_regression
from sklearn.model_selection import train_test_split
from sklearn.model_selection import RandomizedSearchCV
from asgl import Regressor

X, y = make_regression(n_samples=1000, n_features=50, n_informative=25, bias=10, noise=5, random_state=42)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=250)

group_index = np.random.randint(1, 5, size=50)

model = Regressor(model='qr', penalization='asgl', quantile=0.5)

param_grid = {'lambda1': [1e-4, 1e-3, 1e-2, 1e-1, 1], 'alpha': [0, 0.2, 0.4, 0.6, 0.8, 1]}
rscv = RandomizedSearchCV(model, param_grid, scoring='neg_median_absolute_error')
rscv.fit(X_train, y_train, **{'group_index': group_index})

This example demonstrates how to fit a quantile regression model with Adaptive Sparse Group Lasso penalization, utilizing scikit-learn’s RandomizedSearchCV to optimize the model’s hyperparameters.

Example 3: Logistic regression

In binary classification tasks using logistic regression, the default decision threshold of 0.5 is used by default. But it might not always yield the best accuracy. By leveraging the 'logit_proba' model from the asgl package, you can obtain predicted probabilities and use them to find an optimal threshold that maximizes classification accuracy. This example demonstrates how to use cross_val_predict from scikit-learn to evaluate different thresholds and select the one that offers the highest accuracy for your classification model.

import numpy as np
from sklearn.datasets import make_classification
from sklearn.model_selection import train_test_split, cross_val_predict
from sklearn.metrics import accuracy_score, precision_recall_curve
from asgl import Regressor
import matplotlib.pyplot as plt

X, y = make_classification(n_samples=1000, n_features=100, random_state=42)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.25, random_state=42)

# Create a Regressor object for logistic regression to output probabilities
model = Regressor(model='logit_proba', penalization='ridge')

# Use cross_val_predict to get probability estimates for each fold
probabilities = cross_val_predict(model, X_train, y_train, method='predict', cv=5)
#> C:\Users\alvar\ONEDRI~1\Trabajo\Investigacion\asgl\venv\Lib\site-packages\cvxpy\problems\problem.py:1407: UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.
#>   warnings.warn(

thresholds = np.linspace(0.01, 0.99, 100)

# Calculate accuracy for each threshold
accuracies = []
for threshold in thresholds:
    predictions = (probabilities >= threshold).astype(int)
    accuracies.append(accuracy_score(y_train, predictions))

plt.plot(thresholds, accuracies)
plt.title('Accuracy vs Threshold')
plt.ylabel('Accuracy')
plt.xlabel('Threshold')

optimal_threshold = thresholds[np.argmax(accuracies)]
model.fit(X_train, y_train)
test_probabilities = model.predict(X_test)
test_predictions = (test_probabilities >= optimal_threshold).astype(int)
test_accuracy = accuracy_score(y_test, test_predictions)

Example 4: Customizing weights for adaptive sparse group lasso

The asgl package offers several built-in methods for estimating adaptive weights, controlled via the weight_technique parameter. For more details onto the inners of each of these alternatives, refer to the associated research paper or to the user guide. However, for users requiring extensive customization, the package allows for the direct specification of custom weights through the individual_weights and group_weights parameters. This allows the users to implement their own weight computation techniques and use them within the asgl framework.

When using custom weights, ensure that the length of individual_weights matches the number of variables, and the length of group_weights matches the number of groups. Below is an example demonstrating how to fit a model with custom individual and group weights:

import numpy as np
from asgl import Regressor

# Generate custom weights
custom_individual_weights = np.random.rand(X_train.shape[1])
custom_group_weights = np.random.rand(len(np.unique(group_index)))

# Create a Regressor object with custom weights
model = Regressor(model='lm', penalization='asgl', individual_weights=custom_individual_weights, group_weights=custom_group_weights)

# Fit the model
model.fit(X_train, y_train, group_index=group_index)

Example 5: Comparison of lasso and adaptive lasso

This example compares an implementation of lasso as available in scikit-learn against an adaptive lasso model built using the asgl library. Both models are optimized using 5-fold cross validation on a grid of hyper parameters, but ass demonstrated by the final MSEs computed on the test set, the adaptive lasso reduces by half the error compared to lasso.

import numpy as np
from sklearn.linear_model import Lasso
from sklearn.datasets import make_regression
from sklearn.metrics import mean_squared_error
from sklearn.model_selection import GridSearchCV, train_test_split
from asgl import Regressor

X, y = make_regression(n_samples=200, n_features=200, n_informative=25, bias=10, noise=5, random_state=42)

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=50, random_state=42)

param_grid = {'alpha': 10 ** np.arange(-2, 1.51, 0.1)}

lasso_model = Lasso()

gscv_lasso = GridSearchCV(lasso_model, param_grid, scoring='neg_mean_squared_error', cv=5, n_jobs=-1)
gscv_lasso.fit(X_train, y_train)
lasso_predictions = gscv_lasso.predict(X_test)
lasso_mse = np.round(mean_squared_error(lasso_predictions, y_test), 3)
print(f"Lasso MSE: {lasso_mse}")


param_grid = {'lambda1': 10 ** np.arange(-2, 1.51, 0.1)}

alasso_model = Regressor(model='lm', penalization='alasso', weight_technique='lasso')

gscv_alasso = GridSearchCV(alasso_model, param_grid, scoring='neg_mean_squared_error', cv=5, n_jobs=-1)
gscv_alasso.fit(X_train, y_train)
alasso_predictions = gscv_alasso.predict(X_test)
alasso_mse = np.round(mean_squared_error(alasso_predictions, y_test), 3)
print(f"Adaptive lasso MSE: {alasso_mse}")

Lasso MSE: 59.693
Adaptive lasso MSE: 35.085

Contributions

Contributions are welcome! Please submit a pull request or open an issue to discuss your ideas.

Citation

---

If you use asgl in a scientific publication, we would appreciate you cite our paper. Thank you for your support and we hope you find this package useful!

License

This project is licensed under the GPL-3.0 license. This means that the package is open source and that any copy or modification of the original code must also be released under the GPL-3.0 license. In other words, you can take the code, add to it or make major changes, and then openly distribute your version, but not profit from it.