Optional support by XGBoost for easier GPU integration

[DavidHarar]

Hi, First, thank you very much for the great package. Using it lifted my model's KPIs. One thing is, indirectly connected with the miforest, is that it uses light-GBM.
It seems that LGBM has some compatibility issues with GPU and a separate installation (see here). I tried multiple times to install its GPU version without any success, so I ended up using only one or two iterations because of my large dataset. Is there an option to add compatibility with other boosters like XGBoost?

Thank you very much!

[AnotherSamWilson]

Unfortunately no. There is another issue to make the underlying imputation model more modular, but it is a deciptively complex problem. Each model has its own way of dealing with categories, shap values, how the data is encoded, etc. It would be a big project, but not an impossible one.

The underlying model actually used to be sklearn random forests, but they take up a huge amount of memory and are slower. Switching to lightgbm was a big ordeal.

HOWEVER, I would not give up on the lightgbm GPU training. It can be a pain, but what worked for me is installing the wheel version of the package, and simply passing device_type='gpu' into the parameters. I know that lightgbm has a bunch of crazy compilation steps that they show in the tutorial, but apparently the wheel comes pre-built with gpu abilities.

You can verify you are using the gpu version by running a small model in a seperate script with the verbose parameters set to maximum. It will say something like: Using the gpu version of lightgbm! in the output.

[AnotherSamWilson]

@DavidHarar Curious if you got lightgbm up and running with gpu.