LinearRegression on sparse matrices is not sample weight consistent #30131
Labels
Comments
|
Note, that this is not the case for >>> from sklearn.utils.fixes import csr_array
... from sklearn.datasets import make_regression
... from sklearn.linear_model import Ridge
... from sklearn.utils._testing import assert_allclose
...
... X, y = make_regression(100, 100, random_state=42)
... X = csr_array(X)
... reg = Ridge(solver="lsqr", alpha=0, fit_intercept=True, tol=1e-12)
... reg.fit(X, y)
... coef1 = reg.coef_
... reg.fit(X, y, sample_weight=np.ones_like(y))
... coef2 = reg.coef_
... assert_allclose(coef1, coef2, rtol=1e-7, atol=1e-9)But the same fails with a larger >>> reg = Ridge(solver="lsqr", alpha=0, fit_intercept=True, tol=1e-4)
... reg.fit(X, y)
... coef1 = reg.coef_
... reg.fit(X, y, sample_weight=np.ones_like(y))
... coef2 = reg.coef_
... assert_allclose(coef1, coef2, rtol=1e-7, atol=1e-9)
Traceback (most recent call last):
Cell In[9], line 13
assert_allclose(coef1, coef2, rtol=1e-7, atol=1e-9)
File ~/code/scikit-learn/sklearn/utils/_testing.py:232 in assert_allclose
np_assert_allclose(
File ~/miniforge3/envs/dev/lib/python3.12/site-packages/numpy/testing/_private/utils.py:1688 in assert_allclose
assert_array_compare(compare, actual, desired, err_msg=str(err_msg),
File ~/miniforge3/envs/dev/lib/python3.12/contextlib.py:81 in inner
return func(*args, **kwds)
File ~/miniforge3/envs/dev/lib/python3.12/site-packages/numpy/testing/_private/utils.py:889 in assert_array_compare
raise AssertionError(msg)
AssertionError:
Not equal to tolerance rtol=1e-07, atol=1e-09
Mismatched elements: 100 / 100 (100%)
Max absolute difference among violations: 0.00143505
Max relative difference among violations: 0.03345462
ACTUAL: array([ 5.135931e-01, 3.037594e+01, 2.145667e+00, 7.394602e+01,
1.324545e+01, -4.438229e-02, -1.202226e+00, 7.553014e-01,
-1.661572e+00, 1.483041e-01, -3.624484e-01, -1.915631e+00,...
DESIRED: array([ 5.125267e-01, 3.037609e+01, 2.145875e+00, 7.394601e+01,
1.324581e+01, -4.334871e-02, -1.202492e+00, 7.555202e-01,
-1.661108e+00, 1.495079e-01, -3.613935e-01, -1.916570e+00,...I think we need to pass |
16 tasks
1 task
54 tasks
|
I think we should expose |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Part of #16298.
Describe the bug
When using a sparse container like
csr_arrayforX,LinearRegressioneven fails to give the same coefficients for unit or no sample weight, and more generally fails thetest_linear_regression_sample_weight_consitencychecks. In that setting, the underlying solver isscipy.sparse.linalg.lsqr.Steps/Code to Reproduce
Expected Results
The
assert_allcloseshould pass.Actual Results
The test also fails for
fit_intercept=False. Note that this test and other sample weight consistency checks pass if we do not wrapXin a sparse container.Versions
System: python: 3.12.5 | packaged by conda-forge | (main, Aug 8 2024, 18:32:50) [Clang 16.0.6 ] executable: /Users/abaker/miniforge3/envs/sklearn-dev/bin/python machine: macOS-14.5-arm64-arm-64bit Python dependencies: sklearn: 1.6.dev0 pip: 24.2 setuptools: 73.0.1 numpy: 2.1.0 scipy: 1.14.1 Cython: 3.0.11 pandas: 2.2.2 matplotlib: 3.9.2 joblib: 1.4.2 threadpoolctl: 3.5.0 Built with OpenMP: True threadpoolctl info: user_api: blas internal_api: openblas num_threads: 8 prefix: libopenblas filepath: /Users/abaker/miniforge3/envs/sklearn-dev/lib/libopenblas.0.dylib version: 0.3.27 threading_layer: openmp architecture: VORTEX user_api: openmp internal_api: openmp num_threads: 8 prefix: libomp filepath: /Users/abaker/miniforge3/envs/sklearn-dev/lib/libomp.dylib version: NoneEDIT: discovered while working on #30040 for the case of dense inputs.
The text was updated successfully, but these errors were encountered: