Since we did not release, there is no need for a change log entry.

@lesteve I think that this is something like this that you had in mind.

I would argue that the last error reported for Titanic is a bug in LIAC-ARFF. It does not only strip the left and right quotes but the inner quotes. For instance:

"Lovell, Mr. John Hall ('Henry')"

will be parsed as:

"Lovell, Mr. John Hall (Henry)"

while the current PR will not intend to remove the quotes around Henry because it is part of the text.

Edit: Actually this is not the case. I have to check more.
Edit 2: I was stripping all quotes without checking if it was starting and ending with quotes.

I was rumbling in a thread on outdated diff and it is now hidden because of more recent changes. The important comment was:

I am not sure if we should process the values in isolation or based on the column-wide @attribute metadata that should be explicit about the quotation style of the column.

But maybe here we just aim at reproducing the behavior of the old "liac-arff" parser without aiming for correctness and clean data will be fetched by using unambiguous file format such as parquet in the future.

Thinking a but more about this, maybe we could do a study on the available parquet datasets on openml.org to use them as a source of ground truth to make future-proof design decisions on the behavior of the pandas parser: if we don't know if we should automatically strip or not the quotes of nominal variables, maybe we should try to load a few datasets with the pandas / pyarrow integration an look at the values in the resulting pandas dataframe.

Parsing the ARFF attribute metadata should not be difficult if they can help resolve any ambiguity in the CSV payload section of the ARFF file (after the @data line). Feel free to open any dataset from openml.org in a text editor by dowloading the ARFF file by clicking on the Download icon to see what I mean.

Here is the spec:

https://waikato.github.io/weka-wiki/formats_and_processing/arff_stable/

The documentation build passed see https://github.com/scikit-learn/scikit-learn/runs/6808705821?check_suite_focus=true

@thomasjpfan I think this PR is in good shape for final review.

Merging this one, thanks!