Proposal

I'm requesting comment on the following proposal:

• remove keyword argument categorical_feature from cv() and train() in the R and Python packages
• remove keyword argument feature_name from cv() and train() in the Python package
• remove keyword argument colnames from cv() and train() in the R package

And doing all of these only after the packages issuing deprecation warnings for 2-3 releases.

Summary

Both the R and Python packages expose functions cv() (for cross-validation) and train() (for regular entire-dataset training). These functions require a LightGBM Dataset object.

The Dataset object holds attributes categorical_features and feature_names, and allows setting those via constructor keyword arguments and set_{attr}() methods.

Despite that, these cv() and train() functions also take categorical_features and feature_names as keyword arguments.

Python cv()

Python train()

R-package cv()

R-package train()

These keyword arguments aren't providing any value, in my opinion. Their values are just forwarded along to calls like this:

Which at best is redundant with the Dataset class, and at worst could lead to runtime exceptions (if the Dataset has already been constructed).

Motivation

Would simplify the library's interface without any loss of functionality.

If this proposal is accepted, the Dataset class would be the only place that this information is provided to train() and cv().

References

Inspired by this post I noticed on Stack Overflow: https://stackoverflow.com/questions/78383840/in-lightgbm-why-do-the-train-and-the-cv-apis-accept-categorical-feature-argument/78405996#78405996

xgboost does not expose such arguments in train() (code link) or cv() (code link).

These arguments have been part of the API since September 2017: ef77806#diff-9bd633ead0bdfe9540c42a618efd9e559cca16c522ad844a09fcf4ffc7d6e84c.