Tuning LightGBM may appear daunting; it has more than a hundred parameters to tune that you can explore at this page and here.

So many parameters, but honestly, 80% of good performance is the learning rate, and the rest comes from playing with:

• subsampling
• feature_fraction
• min_samples_leaf

And of course the objective function.

Usually, I’ll figure out which learning rate works with default hyperparameters. Then I’ll tweak min_sample_leaf, and then I’ll play around with the rest.

I’ve tried a few methods, but I find you can modify your parameters one by one and it works quite well.

https://twitter.com/marktenenholtz/status/1659920995036307457