Closed
Description
Ridge regression with fit_intercept=True does not give the same result if X is dense or sparse.
The call to _center_data in _BaseRidge.fit should probably be a call to sparse_center_data
test example :
import numpy as np
import scipy.sparse as sp
from sklearn.linear_model import Ridge
from sklearn.utils import safe_sqr
def get_pobj(w, intercept, myX, myy, alpha):
w = w.ravel()
p = np.sum(safe_sqr(myX.dot(w) + intercept - myy)) / 2.
p += alpha * w.dot(w) / 2.
return p
def test_ridge(X, y, alpha):
for solver in ["cholesky", "lsqr", "sparse_cg"]:
clf = Ridge(alpha=alpha, tol=1.0e-15, solver=solver,
fit_intercept=True)
clf.fit(X, y)
print get_pobj(clf.coef_, clf.intercept_, X, y, alpha)
alpha = 1.0
r = np.random.RandomState(42)
X = r.randn(100000, 2)
w = r.randn(2)
i = 10
y = np.dot(X, w) + i
print get_pobj(w, i, X, y, alpha)
print "----Dense----"
test_ridge(X, y, alpha)
print "----Sparse---"
X = sp.csr_matrix(X)
test_ridge(X, y, alpha)returns
1.22411269359
----Dense----
1.22410049215
1.22410049215
1.22410049215
----Sparse---
5.52296274786
5.52296274786
5.52296274786
while with alpha = 0
0.0
----Dense----
4.86608640337e-23
1.41900299631e-26
1.81890174989e-22
----Sparse---
4.2989480748
4.2989480748
4.2989480748