search

skorch-dev / skorch

A scikit-learn compatible neural network library that wraps PyTorch
BSD 3-Clause "New" or "Revised" License
5.88k stars 390 forks source link

Using Skorch NeuralNetClassifier as booster for XGBoost #789

Open francescamanni1989 opened 3 years ago

francescamanni1989 commented 3 years ago

I couldn't re-open issue #737, however do you have any suggestions on how to implement XGBoost by using NeuralNetClassifier as weak learner instead of trees?

Thanks in advance.

BenjaminBossan commented 3 years ago

You probably meant issue #787

For sklearn's GradientBoostingClassifier, it doesn't look like you could easily replace the base estimator, since it is hard-coded: https://github.com/scikit-learn/scikit-learn/blob/2beed55847ee70d363bdbfe14ee4401438fba057/sklearn/ensemble/_gb.py#L195-L207

Theoretically, you could override this method and perhaps be successful, but my guess is that this is going to be very hard since this was not intended by the authors, and it could easily break in new sklearn versions. For XGBoost, this is going to be even more difficult, since significant parts of it are written in C/C++.

I believe your best shot to get something similar would be to use skorch with AdaBoostClassifier as described here.

thomasjpfan commented 3 years ago

There is a discussion in scikit-learn: https://github.com/scikit-learn/scikit-learn/issues/17660 for expanding GradientBoosting* for any base estimator.

Regarding gradient boosting + neural networks, there is recent research on this topic: https://arxiv.org/pdf/2002.07971.pdf (GrowNet).