1
$\begingroup$

heres the link to sklearns feature selection/engineering http://scikit-learn.org/stable/modules/feature_selection.html

I was wondering if these methods are actually useful? I have around 20/30 features (all numerical) in total and was wondering how these methods are able to help?

Shall i perform analysis with and without FS/FE and compare?

$\endgroup$
  • $\begingroup$ Although wording can be improved, the question is just asking if feature selection is helpful in any situation, specially when there are about 20/30 features. A short answer to that question may be useful to future readers. $\endgroup$ – Pere Apr 4 '18 at 7:38
1
$\begingroup$

It really depends on your data, and the amount of data you have.

If you have enough data (intentionally not defining how much "enough" is, cause it really depends on many factors), you can rely on your model to find the best input features for you.

But if for whatever reason you think giving all the features to your model would result in overfitting the model, then you may want to pre-select the features. This may be because you don't have enough data for instance.

If that is the case, then yes, those feature selection methods may be pretty useful, but they're not the only ones you have at hand, and as you see there, they vary a lot and some of them come from training models for the purpose of selecting features, e.g. $L_1$ regularized linear regression, or a tree based feature selection.

You should use your domain knowledge to see which feature selection method would intuitively fit your data, and then use it if necessary.

Also, an easy way is to add some of them to your pipeline, and use a nested cross validation to see if they actually improve your results.

$\endgroup$

Not the answer you're looking for? Browse other questions tagged or ask your own question.