Support forced splits with data and voting parallel versions of LightGBM

[imatiach-msft]

Summary

I'm unable to add forced splits to the data and voting parallel versions of lightgbm in mmlspark, I see the error:

https://github.com/microsoft/LightGBM/blob/master/src/io/config.cpp#L318

Don't support forcedsplits in data tree learne

Motivation

I would like to add this feature but I'm not sure why I can't just remove that line in the config and enable it. What is special about the data and voting parallel learners that wouldn't allow this config to be specified on each node?

Description

The fix would be to remove that thrown exception. Also, it would be great if we could specify the forced splits as a string directly instead of a file.

https://github.com/microsoft/LightGBM/blob/master/src/boosting/gbdt.cpp#L769

[imatiach-msft]

it looks like the ForceSplits function is also only implemented in the serial tree learner:

https://github.com/microsoft/LightGBM/blob/master/src/treelearner/serial_tree_learner.cpp#L450

seems I would need to do something similar in the data parallel tree learner:

https://github.com/microsoft/LightGBM/blob/master/src/treelearner/data_parallel_tree_learner.cpp

and voting parallel learner as well

[shiyu1994]

For data and voting distributed training, we need to synchronize the histogram before GatherInfoForThreshold is called. Otherwise each machine would get a different leaf value and leaf gain, given the incomplete features histograms which sums over only partial data.

LightGBM/src/treelearner/serial_tree_learner.cpp

Lines 483 to 489 in f831808

	leaf_histogram_array[left_inner_feature_index].GatherInfoForThreshold(
	left_leaf_splits->sum_gradients(),
	left_leaf_splits->sum_hessians(),
	left_threshold,
	left_leaf_splits->num_data_in_leaf(),
	left_leaf_splits->weight(),
	&left_split);

So the logic for implementing ForceSplits in data and voting distributed training is not quite straightforward as in feature distributed training.

[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.