[Suggestion] Add power-transforms to AutoMatminer for regression tasks?

hackingmaterials / automatminer

An automatic engine for predicting materials properties.

Other

135 stars 50 forks source link

[Suggestion] Add power-transforms to AutoMatminer for regression tasks? #386

Open CompRhys opened 3 years ago

CompRhys commented 3 years ago

Many ML models perform better when trying to fit approximately Gaussian distributions of values - as such using a power-transform as part of the AMM workflow might lead to a small boost in performance. Not sure if there is a natural way to include into AMM

https://scikit-learn.org/stable/modules/generated/sklearn.preprocessing.PowerTransformer.html

ardunn commented 3 years ago

Hey @CompRhys thanks for the suggestion! Let me look into this some more and get back to you. I don't think this would hurt performance in any way, and could probably be integrated into the functionfeaturizer portion of automatminer. That part of the pipeline is broken right now, but I am planning on fixing it soon. See #217