[MRG + 1] Improve tests for sample_weight in LinearRegression and Ridge by giorgiop · Pull Request #5526 · scikit-learn/scikit-learn

giorgiop · 2015-10-22T11:44:51Z

I have noticed in #5357 that a test introduced in 0.17 was not correctly checking for the effect of sample_weight in LinearRegression. To reproduce the issue on master:

import numpy as np
from sklearn.linear_model.base import LinearRegression

rng = np.random.RandomState(0)
y = rng.randn(6)
X = rng.randn(6, 5)
w = 1.0 + rng.rand(6)

clf = LinearRegression()
clf.fit(X, y, sample_weight=w)
coefs1 = clf.coef_

scaled_y = y * np.sqrt(w)
scaled_X = X * np.sqrt(w)[:, np.newaxis]
clf.fit(scaled_X, scaled_y)
coefs2 = clf.coef_

print(coefs1)
[ 4.58791686 -4.2095038   0.39031788  3.2727146  -0.17386704]
print(coefs2)
[ 3.69763237 -3.64824351  0.367363    2.97550307 -0.44881672]

amueller · 2015-10-22T12:32:18Z

ping @sonnyhu @agramfort @ogrisel ?

amueller · 2015-10-22T12:32:56Z

or @GaelVaroquaux who reviewed the PR together with myself.

amueller · 2015-10-22T12:33:12Z

LGTM

eickenberg · 2015-10-22T12:33:48Z

hmm, shouldn't this also show up in @ainafp 's common tests?

eickenberg · 2015-10-22T12:36:10Z

#5515

sonnyhu · 2015-10-22T13:38:45Z

👍

agramfort · 2015-10-22T13:52:53Z

sklearn/linear_model/base.py

this is not equivalent as now you rescale the non-centered data. Which means that sample weight interact with intercept.

can you clarify why it's more correct this way?

or I saw #5357 (comment)

how to update sci kit to get this bug fixed?

On Thu, Oct 22, 2015 at 9:56 AM, Alexandre Gramfort <
[email protected]> wrote:

In sklearn/linear_model/base.py
#5526 (comment)
:

if sample_weight is not None: # Sample weight can be implemented via a simple rescaling. X, y = _rescale_data(X, y, sample_weight)

X, y, X_mean, y_mean, X_std = self._center_data(

X, y, self.fit_intercept, self.normalize, self.copy_X,

sample_weight=None)

or I saw #5357 (comment)
#5357 (comment)

—
Reply to this email directly or view it on GitHub
https://github.com/scikit-learn/scikit-learn/pull/5526/files#r42750244.

@Sandy4321 we are working on solving the issue at the moment. In the meantime, if you want to use sample weights, you can instead multiply X(row by row) and y with the square root of the weights, as it is done in the code above. Fitting LinearRegression without sample weights still runs without issues.

I can discuss this tomorrow morning IRL

I see , but question is : how technically to do this, simple to use pip to
reinstall?

On Thu, Oct 22, 2015 at 10:42 AM, Giorgio Patrini [email protected]
wrote:

In sklearn/linear_model/base.py
#5526 (comment)
:

if sample_weight is not None: # Sample weight can be implemented via a simple rescaling. X, y = _rescale_data(X, y, sample_weight)

X, y, X_mean, y_mean, X_std = self._center_data(

X, y, self.fit_intercept, self.normalize, self.copy_X,

sample_weight=None)

@Sandy4321 https://github.com/Sandy4321 we are working on solving the
issue at the moment. In the meantime, if you want to use sample weights,
you can instead multiply X(row by row) and y with the square root of the
weights, as it is done in the code above. Fitting LinearRegression
without sample weights still runs without issues.

—
Reply to this email directly or view it on GitHub
https://github.com/scikit-learn/scikit-learn/pull/5526/files#r42756415.

giorgiop · 2015-10-22T14:24:33Z

@agramfort We are reducing a weighted least square problem to a non-weigthed by first performing a change of variables. In the new variables, normalization/centering are weights-independent.

I believe intercept_ has to interact with sample_weight. If we think to the intercept as a dummy column of X, from a linear algebra point of view, a diagonal matrix of the sample weights multiplies that column too.

GaelVaroquaux · 2015-10-22T14:26:21Z

I believe intercept_ has to interact with sample_weight. If we think to the intercept as a dummy column of X, from a linear algebra point of view, a diagonal matrix of the sample weights multiplies that column too.

I am correct in thinking that it's the same picture as a weighted mean?

giorgiop · 2015-10-22T14:34:39Z

It is. We are minimizing a (weighted) L2-norm here. Indeed, the current implementation does the trick a-priori, before the change of variable. See here.

MechCoder · 2015-10-23T06:19:56Z

I think I agree with @giorgiop

More simply, if we just take the loss function as \sum_{i=n} weight * (y - X'w -c)**2 then even the intercept gets weighted by a sqrt(weight) term no?

MechCoder · 2015-10-23T06:47:35Z

sklearn/linear_model/tests/test_base.py

This should just be a big with fit_intercept=True right? But I guess it is good to test.

eickenberg · 2015-10-23T07:09:34Z

iirc when you do the orthogonalization against a constant (ie center), in
order to be able to remove the intercept from the optimization, the
recalculation later on incurs the multiplication of twice that sqrt(w)
factor, so it just ends up being a weighted avg with w.
the second one comes from the fact that we orthogonalize wrt to a weighted
scalar product

On Friday, October 23, 2015, Manoj Kumar [email protected] wrote:

I think I agree with @giorgiop https://github.com/giorgiop

More simply, if we just take the loss function as \sum_{i=n} weight * (y

X'w -c)**2 then even the intercept gets weighted by a sqrt(weight) term
no?

—
Reply to this email directly or view it on GitHub
#5526 (comment)
.

giorgiop · 2015-10-23T07:33:47Z

hmm, shouldn't this also show up in @ainafp 's common tests?

I don't know. I had a look at the test but I don't have an answer just yet.

giorgiop · 2015-10-23T12:21:25Z

So, LinearRegression does not break with the kind of tests of #5515 @eickenberg @ainafb

import numpy as np
from sklearn.linear_model.base import LinearRegression

rng = np.random.RandomState(0)
y = rng.randn(6)
X = rng.randn(6, 5)
w = 2 * np.ones(6)

clf = LinearRegression()
clf.fit(X, y, sample_weight=w)
coefs1 = clf.coef_

scaled_y = np.hstack((y, y))
scaled_X = np.vstack((X, X))
clf.fit(scaled_X, scaled_y)
coefs2 = clf.coef_

coefs1
array([ 4.58791686, -4.2095038 ,  0.39031788,  3.2727146 , -0.17386704])
coefs2
array([ 4.58791686, -4.2095038 ,  0.39031788,  3.2727146 , -0.17386704])

eickenberg · 2015-10-23T12:24:19Z

#5515 shouldn't have a constant sample weight vector. It is drawn uniformly
from a range of integers. That should be different

On Friday, October 23, 2015, Giorgio Patrini [email protected]
wrote:

So, LinearRegression does not break with the kind of tests of #5515
#5515 @eickenberg
https://github.com/eickenberg @ainafb

import numpy as np
from sklearn.linear_model.base import LinearRegression

rng = np.random.RandomState(0)
y = rng.randn(6)
X = rng.randn(6, 5)
w = 2 * np.ones(6)

clf = LinearRegression()
clf.fit(X, y, sample_weight=w)
coefs1 = clf.coef_

scaled_y = np.hstack((y, y))
scaled_X = np.vstack((X, X))
clf.fit(scaled_X, scaled_y)
coefs2 = clf.coef_

coefs1
array([ 4.58791686, -4.2095038 , 0.39031788, 3.2727146 , -0.17386704])
coefs2
array([ 4.58791686, -4.2095038 , 0.39031788, 3.2727146 , -0.17386704])

—
Reply to this email directly or view it on GitHub
#5526 (comment)
.

giorgiop · 2015-10-23T12:32:43Z

Using w = rng.randint(1,5,6) still makes the coefficients equal.
My point is that #5515 cannot cover the issue here and I still have to understand why.

eickenberg · 2015-10-23T13:23:26Z

maybe linear regression is not among the checked estimators for some reason?

On Friday, October 23, 2015, Giorgio Patrini [email protected]
wrote:

My copy and paste mistake. Using w = rng.randint(1,5,6) still makes the
coefficients equal.
My point is that #5515
#5515 cannot cover the
issue here and I still have to understand why.

—
Reply to this email directly or view it on GitHub
#5526 (comment)
.

giorgiop · 2015-11-10T15:59:30Z

@mblondel I think that kind of test is not support to pass. I have removed it and used the idea based on weighted least square that we discussed, with data augmentation by dummy variable. It should be all right. I am amending with some cosmetics as suggest by @agramfort in person, and squashing.

giorgiop · 2015-11-10T16:01:14Z

@agramfort suggested also that we move the test based on weighted least square on test_common, where we may loop over all regressors that minimize square loss, fitting with very small regularization coefficients.

eickenberg · 2015-11-10T16:04:24Z

That looks like it could be done within or in conjunction with the other PR
on a common test for sample weight effects

On Tue, Nov 10, 2015 at 5:01 PM, Giorgio Patrini [email protected]
wrote:

@agramfort https://github.com/agramfort suggested also that we move the
test based on weighted least square on test_common, where we may loop
over all regressors that minimize square loss, which very small
regularization parameters.

—
Reply to this email directly or view it on GitHub
#5526 (comment)
.

giorgiop · 2015-11-10T16:07:45Z

Sure we can. It's going to follow a different logic, but has the same goal.

agramfort · 2015-11-10T16:15:12Z

let me know when I shall review

giorgiop · 2015-11-10T16:18:31Z

let me know when I shall review

Please go ahead!

agramfort · 2015-11-10T20:42:45Z

sklearn/linear_model/tests/test_base.py

coefs3 -> coefs2

agramfort · 2015-11-10T20:45:48Z

also please add note that the same tests are needed for sparse data

agramfort · 2015-11-10T20:45:52Z

that's it for me!

giorgiop · 2015-11-10T21:06:02Z

also please add note that the same tests are needed for sparse data

There is a comment already in the code for this. I will move it to a place where it is more evident.

agramfort · 2015-11-10T21:29:08Z

ok. Let me know when I can do another round

giorgiop · 2015-11-10T21:45:51Z

Comments should have been addressed.

mblondel · 2015-11-10T23:46:32Z

I have removed it and used the idea based on weighted least square that we discussed, with data augmentation by dummy variable.

Indeed this is a much more solid test. Thanks for that!

[MRG + 1] Improve tests for sample_weight in LinearRegression and Ridge

agramfort · 2015-11-11T09:02:02Z

thanks @giorgiop !

amueller · 2015-11-20T22:32:57Z

sklearn/linear_model/tests/test_base.py

this loop is pretty useless now, right?

Yes it is, my bad.

giorgiop changed the title ~~BUG fix weight in LinearRegression~~ BUG fix sample_weight in LinearRegression Oct 22, 2015

giorgiop force-pushed the fix-weight-linearregression branch from b58187f to 66633b8 Compare October 22, 2015 12:19

giorgiop mentioned this pull request Oct 22, 2015

Test sample_weight in common tests #5367

Closed

amueller changed the title ~~BUG fix sample_weight in LinearRegression~~ [MRG] BUG fix sample_weight in LinearRegression Oct 22, 2015

amueller added the Blocker label Oct 22, 2015

amueller added this to the 0.17 milestone Oct 22, 2015

amueller changed the title ~~[MRG] BUG fix sample_weight in LinearRegression~~ [MRG + 1] BUG fix sample_weight in LinearRegression Oct 22, 2015

agramfort reviewed Oct 22, 2015
View reviewed changes

giorgiop force-pushed the fix-weight-linearregression branch 2 times, most recently from 70ea898 to c1c35d0 Compare October 22, 2015 14:42

MechCoder reviewed Oct 23, 2015
View reviewed changes

giorgiop force-pushed the fix-weight-linearregression branch 2 times, most recently from 2c2897a to c7e276a Compare October 23, 2015 10:07

giorgiop force-pushed the fix-weight-linearregression branch 2 times, most recently from b2e6dcb to 2dcad5e Compare November 10, 2015 15:59

giorgiop force-pushed the fix-weight-linearregression branch from 2dcad5e to b76d5a4 Compare November 10, 2015 16:07

giorgiop changed the title ~~[MRG + 1] BUG fix sample_weight in LinearRegression~~ [MRG + 1] Improve tests for sample_weight in LinearRegression and Ridge Nov 10, 2015

agramfort reviewed Nov 10, 2015
View reviewed changes

sklearn/linear_model/tests/test_base.py Outdated

Copy link
Copy Markdown

Member

agramfort Nov 10, 2015

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

coefs3 -> coefs2

tests sample_weights linearreg, ridge

1d16ec4

giorgiop force-pushed the fix-weight-linearregression branch from b76d5a4 to 1d16ec4 Compare November 10, 2015 21:08

agramfort added a commit that referenced this pull request Nov 11, 2015

Merge pull request #5526 from giorgiop/fix-weight-linearregression

a5d6144

[MRG + 1] Improve tests for sample_weight in LinearRegression and Ridge

agramfort merged commit a5d6144 into scikit-learn:master Nov 11, 2015

giorgiop deleted the fix-weight-linearregression branch November 11, 2015 17:40

giorgiop mentioned this pull request Nov 12, 2015

[WIP] Sample weight consistency #5515

Closed

amueller reviewed Nov 20, 2015
View reviewed changes

Uh oh!

Conversation

giorgiop commented Oct 22, 2015

Uh oh!

amueller commented Oct 22, 2015

Uh oh!

amueller commented Oct 22, 2015

Uh oh!

amueller commented Oct 22, 2015

Uh oh!

eickenberg commented Oct 22, 2015

Uh oh!

eickenberg commented Oct 22, 2015

Uh oh!

sonnyhu commented Oct 22, 2015

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

agramfort Oct 22, 2015 via email

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

giorgiop commented Oct 22, 2015

Uh oh!

GaelVaroquaux commented Oct 22, 2015 via email

Uh oh!

giorgiop commented Oct 22, 2015

Uh oh!

MechCoder commented Oct 23, 2015

Uh oh!

Choose a reason for hiding this comment

Uh oh!

eickenberg commented Oct 23, 2015

Uh oh!

giorgiop commented Oct 23, 2015

Uh oh!

giorgiop commented Oct 23, 2015

Uh oh!

eickenberg commented Oct 23, 2015

Uh oh!

giorgiop commented Oct 23, 2015

Uh oh!

eickenberg commented Oct 23, 2015

Uh oh!

giorgiop commented Nov 10, 2015

Uh oh!

giorgiop commented Nov 10, 2015

Uh oh!

eickenberg commented Nov 10, 2015

Uh oh!

giorgiop commented Nov 10, 2015

Uh oh!

agramfort commented Nov 10, 2015 via email

Uh oh!

giorgiop commented Nov 10, 2015

Uh oh!

Choose a reason for hiding this comment

Uh oh!

agramfort commented Nov 10, 2015

Uh oh!

agramfort commented Nov 10, 2015

Uh oh!

giorgiop commented Nov 10, 2015

Uh oh!

agramfort commented Nov 10, 2015 via email

Uh oh!

giorgiop commented Nov 10, 2015

Uh oh!

mblondel commented Nov 10, 2015

Uh oh!

agramfort commented Nov 11, 2015

Uh oh!

Choose a reason for hiding this comment