Choosing good classifiers for forward selection applied on nominal data

green_tea · January 2019

Hello community,

my goal is to run a wrapper-based feature selection on ~70 nominal features to select a the ~10 best ones. I think a forward selection is the best choice here as it starts with no features and adds one new feature at a time. I read through several guides here on how to do a wrapper-based feature selection that were very helpful in implementing this.

However I am still lost on which classifiers I should select inside the model. I will not use the resulting dataset to train and test a model afterwards, so the obvious choice of selecting the same classifier as I would for the model is not there. Are there any posts here I missed so far that would help me with selecting classifiers? Or can you share your knowledge and experience on this with me? I greatly appreciate your answers!

MartinLiebig · January 2019

Hi,

i am not sure what the best selection is, but Naive Bayes should be in. At least as a base line.

BR,

Martin

lionelderkrikor · January 2019

Hi @green_tea,

Don't focus only on the "forward selection".
I think that in life (and thus in data-science), it is always relevant to compare.
RapidMiner propose several methods of feature selection :

Image: https://us.v-cdn.net/6030995/uploads/editor/nr/dldp0bsi1ljd.png

You can test them.
From my own experience, the Optimize Selection (Evolutionary) operator gave always me good results...

To conclude, here a link to a ressource relativ to Feature Selection :
https://community.rapidminer.com/discussion/45775/multi-objective-feature-selection-part-1-the-basics

Hope it helps,

Regards,

Lionel