[R-package] Accept data frames as inputs

[jameslamb]

Summary

It should be possible to use R data frames directly as inputs in {lightgbm}, without converting them to matrices.

Motivation

The data.frame is a very common data structure in R, and many statistics and machine learning projects accept data frames as inputs, including but not limited to {caret}, {randomForest}, and the packages listed in #4295 (comment).

Allowing the use of data.frame objects would reduce friction for R users working with LightGBM.

Description

I think this work can be broken down into the following components

• accept data.frame inputs for data argument to lgb.Dataset()
• accept data.frame inputs for data argument to lightgbm()
• accept data.frame inputs for data argument to lgb.cv()
• automatically encode categorical-looking column types (factor, character) in a LightGBM-friendly way
• accept data.frame inputs for predict.lgb.Booster() / Booster$predict() / Predictor$predict()
  • once "automatically encode categorical-looking column types" is implemented, this should also including converting categorical-looking columns with the same rules
• allow passing a column name for init_score in lgb.Dataset() if data is a data.frame
• allow passing a column name for label in lgb.Dataset() if data is a data.frame
• allow passing a column name for weight in lgb.Dataset() if data is a data.frame

References

See #4207 for a prior proposal and #4207 (review) for reference on this issue.

This feature precedes #4295.

[jameslamb]

I've added this feature request to #2302, per this project's standard approach for managing its backlog.

If you are interested in contributing to this feature or have additional information to add, please leave a comment and the issue can be re-opened.

[jameslamb]

Want to add that when this feature is picked up, the R package should support categorical features for data frames using the same interface as the Python package:

• setting categorical_feature = "auto" in Dataset (the default) means "automatically detect categorical features"
• if categorical_feature in Dataset is a vector of feature indices or names, use that list of features for categorical_feature instead

LightGBM/python-package/lightgbm/basic.py

Line 1127 in 8a90ea3

feature_name='auto', categorical_feature='auto', params=None,

LightGBM/python-package/lightgbm/basic.py

Lines 534 to 540 in 8a90ea3

	if categorical_feature is not None:
	if feature_name is None:
	feature_name = list(data.columns)
	if categorical_feature == 'auto': # use cat cols from DataFrame
	categorical_feature = cat_cols_not_ordered
	else: # use cat cols specified by user
	categorical_feature = list(categorical_feature)

LightGBM/python-package/lightgbm/basic.py

Lines 555 to 556 in 8a90ea3

	if categorical_feature == 'auto':
	categorical_feature = None

[mayer79]

The tricky thing here is that we would need to store factor levels in the Booster object in order to allow for safe out-of-sample predictions. This could be a named list like xlevs = list(gender = c("m", "f"), age = c("young", "old")).

[jameslamb]

Sorry, this was locked accidentally. Just unlocked it. We'd still love help with this feature!