PCA.transform fails if svd_solver="randomized"

[hanslovsky]

Describe the issue:

dask_ml.decomposition.PCA.transform fails after running dask_ml.decomposition.PCA.fit for svd_solver="randomized":

AttributeError: 'PCA' object has no attribute 'power_iteration_normalizer'

This does not happen for dask_ml.decomposition.PCA.fit_transform

Minimal Complete Verifiable Example:

# /// script
# requires-python = "==3.14"
# dependencies = [
#   "dask-ml==2025.1.0",
#   "scikit-learn==1.8.0",
# ]
# ///

import dask.array as da
from dask_ml.decomposition import PCA
from sklearn.datasets import load_breast_cancer

data = da.asarray(load_breast_cancer().data)
solver = "randomized"

pca1 = PCA(svd_solver=solver)
data1 = pca1.fit_transform(data)
print("fit_transform worked")

pca2 = PCA(svd_solver=solver)
pca2.fit(data)
data2 = pca2.transform(data)
print("fit, then transform worked")

Assuming this snippet is saved as mwe.py, run with uv:

uv run mwe.py

The expected output is:

fit_transform worked
Traceback (most recent call last):
  File "/home/zottel/workspace/dask-ml-pca/mwe.py", line 22, in <module>
    data2 = pca2.transform(data)
  File "/home/zottel/.cache/uv/environments-v2/mwe-fa30d1c07b89220f/lib/python3.14/site-packages/sklearn/utils/_set_output.py", line 316, in wrapped
    data_to_wrap = f(self, X, *args, **kwargs)
  File "/home/zottel/.cache/uv/environments-v2/mwe-fa30d1c07b89220f/lib/python3.14/site-packages/dask_ml/decomposition/pca.py", line 373, in transform
    check_is_fitted(self, ["mean_", "components_"])
    ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/zottel/.cache/uv/environments-v2/mwe-fa30d1c07b89220f/lib/python3.14/site-packages/sklearn/utils/validation.py", line 1699, in check_is_fitted
    tags = get_tags(estimator)
  File "/home/zottel/.cache/uv/environments-v2/mwe-fa30d1c07b89220f/lib/python3.14/site-packages/sklearn/utils/_tags.py", line 275, in get_tags
    tags = estimator.__sklearn_tags__()
  File "/home/zottel/.cache/uv/environments-v2/mwe-fa30d1c07b89220f/lib/python3.14/site-packages/sklearn/decomposition/_pca.py", line 845, in __sklearn_tags__
    solver == "randomized" and self.power_iteration_normalizer == "QR"
                               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
AttributeError: 'PCA' object has no attribute 'power_iteration_normalizer'

Anything else we need to know?:
I was able to reproduce this error with Python versions 3.12, 3.13, and 3.14.

Environment:

• Dask version: 1.8.0
• Python version: 3.12, 3.13, 3.14
• Operating System: Linux (but reproduced on macOS as well)
• Install method: uv from pypi.org through inline script dependencies defined in the snippet above.

[hanslovsky]

The error happens in the check_is_fitted call in PCA.transform. Ultimately, the get_tags function calls into sklearn's PCA.__sklearn_tags__ and fails when ]checking for array_api_support](https://github.com/scikit-learn/scikit-learn/blob/1.8.0/sklearn/decomposition/_pca.py#L844-L846):

        tags.array_api_support = solver not in ["arpack", "randomized"] or (
            solver == "randomized" and self.power_iteration_normalizer == "QR"
        )

I don't know what the best solution would be. Two options come to mind:

1. Override __sklearn_tags__ without calling tags = super().__sklearn_tags__(). This probably is not a great option because the base estimator may do some important things here. There is a probably a way to call sklearn PCA's __sklearn_tags__ without calling __sklearn_tags__ for sklearn PCA.
2. Set power_iteration_normalizer=None. This is also odd because we set a variable that we probably never use.

Not super happy with either of those solutions. Better options probably require a bit of a deeper dive and/or better understanding of how ml-dask PCA and sklearn (version 1.8.0) PCA differ.

For now, I can just use PCA.fit_transform as a workaround. That should be ok for my use case.