sparsereg.vendor.group_lasso package¶
Submodules¶
sparsereg.vendor.group_lasso.group_lasso module¶
-
sparsereg.vendor.group_lasso.group_lasso.
sparse_group_lasso
(X, y, alpha, rho, groups, max_iter=1000, rtol=1e-06, verbose=False)[source]¶ Linear least-squares with l2/l1 + l1 regularization solver.
Solves problem of the form:
- (1 / (2 n_samples)) * ||Xb - y||^2_2 +
- [ (alpha * (1 - rho) * sum(sqrt(#j) * ||b_j||_2) + alpha * rho ||b_j||_1) ]
where b_j is the coefficients of b in the j-th group. Also known as the
sparse group lasso
.Parameters: - X (array of shape (n_samples, n_features)) – Design Matrix.
- y (array of shape (n_samples,)) –
- alpha (float or array) – Amount of penalization to use.
- groups (array of shape (n_features,)) – Group label. For each column, it indicates its group apertenance.
- rtol (float) – Relative tolerance. ensures ||(x - x_) / x_|| < rtol, where x_ is the approximate solution and x is the true solution. TODO duality gap
Returns: x – vector of coefficients
Return type: array
References
“A sparse-group lasso”, Noah Simon et al.