Hi, I understand that this may look llike an expected behaviour but this can lead to unexpected results in the following scenario:

• arff file with quoted question marks as categorical values and data: e.g. @attribute feat1 {'?', 'A', 'B', 'C'}
• arff.load() reads those '?' as strings.
• arff.write() (for example after sampling the original data) then writes the '?' from loaded data without quotes: @attribute feat1 {?, A, B, C}
• arff.load() the last file interpretes ? as missing value (None).

see openml/automlbenchmark#209 for a hack implemented locally to prevent this, but this hack also means that it would not be possible anymore to represent missing values as ? in arff files saved with the library.

Suugesting to add a param to arff.dump signature, for example:

def dump(obj, fp, missing_values=[None, '?']):
    pass

allowing user to call arff.dump(o, f, missing_values=[None]) when ? should not be considered as a missing value, and therefore be quoted.