LightGBM overestimating non-zero-classes
I have a classification problem with 3 classes. The signal-to-noise ratio is rather low, and the non-zero classes are about 3-4% of all data. So I decided to weigh the samples inversely by their frequency:
LightGBM – how many trees do I actually have?
Beginner experimenting with LGBM here. I have code that looks like this