Repository navigation
[ENH] Automatically convert non-string categorical data in attributes_arff_from_df #1489
Description
Is your feature request related to a problem? Please describe.
Currently, attributes_arff_from_df raises a ValueError if a pandas DataFrame contains a categorical column with non-string values (e.g., integers [0, 1]). The user is forced to manually cast these to strings before passing the DataFrame.
Describe the solution you'd like
Instead of raising an error immediately, the function should attempt to automatically convert the categories to strings using .astype(str). This improves UX for users working with mixed-type or integer-encoded categorical data.
Describe alternatives you've considered
Keep raising the error, but improve the message. However, automatic conversion is more user-friendly as ARFF expects string nominals anyway.
Additional context
I have a fix implemented locally and can submit a PR.
Activity
[ENH] Auto-convert categorical columns to string in attributes_arff_f…
alphaleporus commentedon Nov 20, 2025
PR submitted!
AtharvaManale commentedon Oct 13, 2026
Hi @PGijsbers,
I have reviewed the issue and the proposed automatic conversion of non-string categorical labels for #1489, I explored a couple of edge cases and added tests for integer categories and categorical columns containing missing values which the proposed solution by @alphaleporus failed.
I have raised a pr #1771 which showcases the testcases and what problems the proposed solution creates further.
Would you recommend keeping the current ValueError behavior for unsupported categorical types, or should we pursue automatic conversion with an explicitly defined policy?
If the current behavior of returning valueerror is intentional, perhaps the issue could be closed or reframed as an enhancement. I’d appreciate your guidance on the issue and the existing PR.