[python-package] Interpreting the contents of dump_model for multi-label classification

[Opt-Mucca]

I am currently having issues determining which trees correspond to which class from the output of LGBMClassifier.booster_.dump_model().

Take a simple scenario of LGBMClassifier has n_classes_=3 and n_estimators_=2. From my understanding of GBDTs and RFs in LightGBM, for each class label n_estimators_ are constructed. This is supported by the output of raw = LGBMClassifier.booster_.dump_model(), where raw["tree_info"] is a list containing 2*3 decision trees.

My question is:
How do I know which set of n_estimators_ decision trees correspond to a given class?

In the output of raw = LGBMClassifier.booster_.dump_model() I can find no information on which class from the set {0, 1, 2} raw["tree_info"][i] corresponds to for i in {0,...,5}. So let the classes be {0, 1, 2}. Are the 6 trees assumed to be ordered in any of the following ways?: [0, 1, 2, 0, 1, 2] or [0, 0, 1, 1, 2, 2]. If not, where would I find this information? (I have looked into the reader and believe it is doing the first option [0, 1, 2, 0, 1, 2] , but would like some confirmation. I would also assume this information should be somewhere in the output of dump_model, but I cannot find anything).

[jameslamb]

Thanks for using LightGBM.

for each class label n_estimators_ are constructed

This is correct specifically in the case of multi-class classification.

For binary classification, there will be n_estimators_ trees (assuming no early stopping).

are the 6 trees assumed to be ordered in any of the following ways?

Using the terminology from your example, the trees for a multiclass classification model are ordered [0, 1, 2, 0, 1, 2, ...].

You could also confirm this by checking the predictions, e.g. like this example with the R package: #5223 (comment).

[Opt-Mucca]

Thanks for the great response! It perfectly answers the question. Additional thanks for even specifying the binary case.

[github-actions]

This issue has been automatically locked since there has not been any recent activity since it was closed. To start a new related discussion, open a new issue at https://github.com/microsoft/LightGBM/issues including a reference to this.