Why train model on smaller and smaller set of features recursively about featurewiz HOT 3 CLOSED

Great question @MySciencePlayground 👍
The reason I chose this method is that I found permutation based methods where:

1. Multiple permutations of features are selected sequentially is very time consuming and for large datasets, can be explosive in terms of cost
2. Random selection of baskets of features may leave out certain features and that could be costly.
   The best compromise between the two methods is this sequential selection method where smaller baskets of features are selected and it scales well even when 100's or 1000's of features are in a dataset. That's why it works well for large datasets as well as for small datasets.
   Hope this clarifies
   AutoVimal

from featurewiz.

Thanks for the quick respond, @AutoViML! So with the current approach, the column order can have some effect on the final selection. Out of curiosity, is there any test/evaluation on how different the results could be ? (applying same dataset but changing the column order)

from featurewiz.

Hi @MySciencePlayground
That's a good question but I was hoping you would try it yourself out of curiosity by changing the column order and testing it. I have done my tests and I will leave it to you to try it and report back your findings and questions if any 👍

Please upgrade to the latest version since it fixes a bug related to this issue:

 !pip install featurewiz --ignore-installed --no-deps --upgrade

This is the tryout medal for your efforts!🥇

AutoVimal

from featurewiz.