Support ignoring some features during training on constructed dataset

[wangmn93]

I want to get a subset of features from a contructed dataset because reconstructing the dataset is time-consuming. Is there any way to do this or any suggestion on how to modify the source code?

Another way is ignoring some features on constructed dataset, but this feature does not work on constructed dataset.

[shiyu1994]

@wangmn93 Thanks for using LightGBM. Currently LightGBM don't support extracting a subset of features from a constructed data. But I think there's a quick hack to do this.

In addition to ignore_column, which is used when constructing dataset, we can add a new parameter ignore_column_training, which is dedicated to ignore some features during training. Then we can parse ignore_column_training in the same way as ignore_column, and store it in the config_ object. Then in src/treelearner/serial_tree_learner.cpp, we can set the is_feature_used according to ignore_column_training in the SerialTreeLearner::FindBestSplits method.

In this way, ignore_column_training is not involved in the dataset constructing process. Instead, it only affects the training process, and can be safely changed through setting params in the lgb.train method.

If you have any further problem about the implementation, or need any other help, please feel free to post here.

[shiyu1994]

BTW, when setting is_feature_used, we should notice that is_feature_used uses the so called inner feature index of LightGBM, which is different from the real feature index in the input data. So we need to remap the real feature index specified by ignore_column_training to inner feature index through train_data_->InnerFeatureIndex method.

[StrikerRUS]

Closed in favor of being in #2302. We decided to keep all feature requests in one place.

Welcome to contribute this feature! Please re-open this issue (or post a comment if you are not a topic starter) if you are actively working on implementing this feature.