Learning with soft labels

[sinchir0]

Summary

I would like LightGBM to be able to learn with the following soft labels.

example

y_train = np.array(
    [
        [0.3, 0.3, 0.4],
        [0.1, 0.7, 0.2],
        [0.2, 0.3, 0.5]
    ]
)

If the number of classes is 3, it is not impossible to use the custom objective function and copy the data 3 times to calculate loss. However, if the data is huge, it will consume a lot of memory and take a long time to calculate.

Motivation

For example, for a multiclass task, it is possible to have more detailed information in the objective.