DOC Fix Wikipedia principal eigenvector example - #34266

I guess @ogrisel already has a lot on his plate, maybe we can just look for someone else's review?

Out of interest, @ArturoAmorQ how did you bump into the issue? Do you run each morning a random file from the scikit-learn example gallery 😜? Maybe you used a LLM to try to find examples to improve 🤖?

I pushed a tiny tweak and since the example wasn't taking that much time locally (1min30s) I thought it could be run as part of the doc build so I renamed it to add plot_ which automatically will make it run in the doc build (that's the default convention in sphinx-gallery IIRC).

Not sure what happened with the CI linting issue hoping merging main will fix it ...

For completeness, the error was:

### Running cython-lint ###

/home/circleci/project/sklearn/cluster/_dbscan_inner.pyx:31:25: Outer for loop variable 'i' overwritten by inner for-loop target
/home/circleci/project/sklearn/cluster/_k_means_common.pyx:30:9: Loop control variable 'i' not used within the loop body (if this is intended, start the name with an underscore)
/home/circleci/project/sklearn/manifold/_utils.pyx:74:13: Loop control variable 'l' not used within the loop body (if this is intended, start the name with an underscore)
Problems detected by cython-lint, please fix them

Out of interest, @ArturoAmorQ how did you bump into the issue? Do you run each morning a random file from the scikit-learn example gallery 😜? Maybe you used a LLM to try to find examples to improve 🤖?

I bumped into it when experimenting methodologies to rank the recommended examples in our gallery. My initial idea was that more connected examples are more relevant.

I pushed a tiny tweak and since the example wasn't taking that much time locally (1min30s) I thought it could be run as part of the doc build so I renamed it to add plot_ which automatically will make it run in the doc build (that's the default convention in sphinx-gallery IIRC).

It's mostly the first download of the data that is slow.

Looking at the CircleCI log, the example when downloading the dataset takes less than 3 minutes which isn't that bad:

    - ../examples/applications/plot_wikipedia_principal_eigenvector.py:                             165.91 sec   0.0 MB

The problem is more the size of the downloaded files that takes 900MB:

❯ du -sh ~/scikit_learn_data/downloads.dbpedia.org/3_5_1_en/*
835M	/home/lesteve/scikit_learn_data/downloads.dbpedia.org/3_5_1_en/page_links_en.nt.bz2
57M	/home/lesteve/scikit_learn_data/downloads.dbpedia.org/3_5_1_en/redirects_en.nt.bz2

That would make the Circle cache at 2-3 times bigger for just a single example, making it slower to upload/download in CircleCI, so in the end I am not sure it's worth it. Also people that run all the examples (not that many but maybe some sprint contributors are instructed to do make html ¹) would end up with 900MB more of occupied space, which can be surprising ...

I'll wait for the CircleCI doc to finish, maybe push another commit to see the impact on CircleCI cache and decide whether we rename the example back to what it was i.e. without plot_ so that it isn't run in the doc build.

Funnily enough this reminded me of one of my old issue (super-low priority) #8849, where running the Wikipedia example as part of the CI was mentioned 😉.

Looking at the logs the cache doesn't take much time to save/restore in CircleCI (less than seconds), I propose we run it as part of the CI.

The slight downside is that someone doing make html will download an additional 900MB, I think this is acceptable. If somebody disagrees we can always revert the name change to wikipedia_principal_eigenvector.py (i.e. no plot_ at the beginning).

The plot looks fine I have enabled auto-merge: https://output.circle-artifacts.com/output/job/af5ae68e-991d-49b2-8002-be58315c07cf/artifacts/0/doc/auto_examples/applications/plot_wikipedia_principal_eigenvector.html#plot-results