ignore_column seems obsolete

[kainkad]

Description

All of the Dataset parameters seem to work correctly, except for ignore_column https://lightgbm.readthedocs.io/en/latest/Parameters.html#ignore_column.
I have seen previous similar issues on this topic which are quite old and closed such as #1061, #2376, or #4657 so I understand why it wouldn't be working for a binary file that is already a constructed dataset, but why wouldn't it be working for a csv file? I suppose there are other ways of blacklisting or dropping columns from the dataset rather than using the lightgbm Dataset parameters but since it is there and seems available, I would like to reconfirm that is working as expected and that perhaps my reproducible example is not correct.

Reproducible example

import lightgbm as lgb
import pandas as pd

# Create a DataFrame with sample data
sample_data = pd.DataFrame({
    'label': [0, 1, 0, 1, 0, 1],
    'feature2': [5, 7, 2, 6, 1, 8],
    'feature3': [3, 8, 3, 7, 2, 9],
    'feature4': [4, 9, 4, 8, 3, 10],
    'group_id': [1, 2, 2, 3, 3, 3]
})
# Save the DataFrame to a CSV file
sample_file = 'sample_data.csv'
sample_data.to_csv(sample_file, index=False)

# Load the dataset with LightGBM (label is by default the first column but making it explicit)
dataset = lgb.Dataset(sample_file, params={'header': True, 'label_column': 'name:label', 'ignore_column': 'name:feature2,feature3', 'query': 'name:group_id'})

# Construct the dataset
dataset.construct()

# [LightGBM] [Info] Using column label as label
# [LightGBM] [Info] Using column group_id as group/query id
# [LightGBM] [Warning] There are no meaningful features, as all feature values are constant.
# [LightGBM] [Info] Construct bin mappers from text data time 0.00 seconds

# Get the number of features and data points
num_feature = dataset.num_feature()
print(f'Number of features: {num_feature}') # shows 4 incorrectly?
num_data = dataset.num_data() # shows 6 correctly
print(f'Number of data points: {num_data}')
# the two below work correctly (label is by default the first column)
get_label = dataset.get_label()
print(f'Label: {get_label}')
get_group = dataset.get_group()
print(f'Group: {get_group}')

Environment info

Operating System: Windows 11

CPU/GPU model: CPU

C++/Python/R version: Python

LightGBM version or commit hash: 4.5.0 (same issue with v3)

Given the above sample data, I would expect the number of features to be only the feature4 with the label and group_id columns not being treated as features and the feature2 and feature3 being ignored/dropped. However, the number of features returned is 4 (which I understand only ignores the label column and loads the remaining 4 columns). I also tried with using the number of indices instead of the name: and using a plain text CSV read as csv instead of using a dataframe. I also thought that may be .num_features() can be retrieved only before constructing a dataset but the features cannot be obtained before constructing the dataset as the clear error message is returned in this case.
I also thought that may be this parameter only works in CLI as it was suggested in one of the older issues but then the documentation does not state that.

[kainkad]

Adding to the above if I set the ignore_column parameter to contain indices, or names that don't exist in the data, the number of features still remains the same, e.g.

dataset = lgb.Dataset(sample_file, params={'header': True, 'label_column': 'name:label', 'ignore_column': '0,1,2,3,4,5,6,7,8,9,10', 'query': 'name:group_id'})
dataset.construct()
num_feature = dataset.num_feature() # gives 4 as well