[ENH] Add encoding support for categorical variables in v2

[PranavBhatP]

Is your feature request related to a problem? Please describe.
Currently encoding support is missing for categorical variables in v2 of PTF. Efforts are being made to implement scaling in #1983. Feature encoding is naturally the next step after this. Re also: #1993

Describe the solution you'd like
Implement a robust encoding pipeline for handling categorical variables in the TimeSeries dataset.

[ishan372or]

Hi, I’d like to work on implementing a robust encoding pipeline for categorical variables in TimeSeries datasets for PTF v2. Could I be assigned this issue?

[PranavBhatP]

Hi @ishan372or, welcome! If you are a new contributor, feel free to look at other good first issues as well. I would say that this issue is somewhat of a challenge and can have a long review cycle since certain parts of design has not been fully decided with respect to scaling and encoding, so it might not be the most convenient one (if u are getting started). Otherwise, if you want take a shot at this, you can go ahead of course! I will assign you, in that case.

[ishan372or]

Hi @PranavBhatP , thanks for the clarification! I understand that the design is still evolving and that this may take longer to review. I’m okay with that and would like to take a shot at it. I’ll start with a minimal, well-scoped implementation and iterate based on feedback.

[ishan372or]

Hi @PranavBhatP,Currently, categorical features are only handled via NaNLabelEncoder, which provides basic label encoding but is not integrated into a dataset-level pipeline for TimeSeries. For an initial implementation, I propose adding a lightweight dataset-level categorical encoding step that reuses NaNLabelEncoder for ordinal encoding, supports explicitly user-specified categorical columns, and fits encoders on training data for reuse in validation/test. The initial scope would be limited to static categorical covariates and ordinal encoding only, with one-hot or embedding-based encoding, dynamic categorical features, automatic dtype inference, and tight coupling with scaling/normalization considered out of scope. Happy to adjust the scope or design based on feedback.

[PranavBhatP]

Sounds good!