kaikaiguo/smoothed_concomitant_lasso

Smooth Concomitant Lasso

This package implements a simple scikit-learn style estimator for the Smoothed Concomitant Lasso https://arxiv.org/abs/1606.02702.

Motivation

Automated noise level estimatotion can be of interest in high dimension, for instance when uncertainty quantification is required, as well as when no cross-validation can be performed to tune the regularization paramter.

Example

The API of the Smoothed_Concomitant_Lasso code is designed for compatibility with the scikit-learn project.

import numpy as np
import matplotlib.pyplot as plt
from smoothed_concomitant import SC_path
from SBvG import SBvG_path
from data_generation import generate_data

# Generate dataset

n_samples = 100
n_features = 100
sigma = 2
sparsity = 0.9
snr = 1
correlation = 0.5
random_state = 42

X, y, true_beta, true_sigma = generate_data(n_samples, n_features, sigma, snr,
                                            sparsity, correlation,
                                            random_state=random_state)

# regularization parameter for sigma estimation
sigma_0 = (np.linalg.norm(y) / np.sqrt(n_samples)) * 1e-2
sigstar = max(sigma_0, np.linalg.norm(y) / np.sqrt(n_samples))
lambda_max = np.linalg.norm(np.dot(X.T, y), ord=np.inf) / (n_samples * sigstar)

# SC
betas, sigmas, gaps, n_iters = SC_path(X, y, [lambda_max / 1.5],
                                       eps=1e-4, max_iter=5000)

# SBvG
betas_SBvG, sigmas_SBvG = SBvG_path(X, y, [lambda_max / 1.5])


fig, axes = plt.subplots(ncols=3, sharey=True)

axes[0].stem(true_beta)
axes[0].set_title("True signal")
axes[1].stem(betas[0])
axes[1].set_title("SC")
axes[2].stem(betas_SBvG[0])
axes[2].set_title("SBvG")

plt.show()


plt.savefig("test_SC.png", format="png")

# Sigma performance:
print("Sigma estimation:")
print("True sigma")
print(sigma)

print "hatsigma for SC"
print(sigmas[0])

print "hatsigma for SBvG"
print(sigmas_SBvG[0])

# Beta performance
print("Estimation risk")
print("SC")
print(np.linalg.norm(true_beta - betas[0]) / n_features)
print("SBvG")
print(np.linalg.norm(true_beta - betas_SBvG[0]) / n_features)

Installation & Requirements

The Smoothed_Concomitant_Lasso package itself is fairly lightweight. It is tested on Python 2.7 and 3.4-3.5, and depends on the following packages:

• numpy
• scipy
• scikit-learn
• cython

Using the cross-platform conda package manager, these requirements can be installed as follows:

$ conda install numpy scipy scikit-learn cython

To install Smoothed_Concomitant_Lasso from source, first download the source repository and then run

$ cython cd_smoothed_concomitant.pyx
$ cython cd_smoothed_concomitant_screening.pyx
$ python setup.py build_ext --inplace