LightGBM: Support for outputting decision path and leaf

[ajryan]

Has any work already begun for outputting decision path and leaf data? I believe it would be similar to this implementation for decision tree and random forest in sklearn_onnx onnx/sklearn-onnx#563.

One challenge appears to be onnxmltools does not have a story for converters that accept options. I see that sklearn implemented their own topology and Scope that exposes an options API.

[ajryan]

Started work on this here: https://github.com/ajryan/onnxmltools/tree/aidanryan/feature/decision-path

[xadupre]

I did that work for models in scikit-learn: https://onnx.ai/sklearn-onnx/auto_tutorial/plot_dbegin_options.html#option-decision-path but I did not do it for xgboost or lightgbm. The way to do it is to generate a second tree producing the leave id as a target, then link that id with the value you need to produce for that leave with a LabelEncoder.

[akhiliyengar]

The current implementation for Random forest uses a regressor for each tree. We can achieve this using a single regressor and map the weights to the leaf id using a single regressor.