1

I recently came across an approach where the long list of potential predictors (around 100+) is screened for its explanatory power and basis this initial screening, a smaller set of predictors is arrived at (which is then used for model building) The screening technique adopted is the "Information Value" approach (There has been a query on Information value approach before - Why do we calculate Information value?. And one of the users have provided an easy to understand explanation of the same)

However, my question is, in what scenarios would you like to do an initial screening of variables prior to building a model (considering that the model building itself involves techniques for variable selection)? Also, if the issue is with the number of predictors, won't a dimensionality reduction algorithm like PCA suffice? So, in nutshell, what is the value that an initial screening approach like IV provide?

Thanks!

Community
  • 1
Dataminer
  • 365
  • 3
  • 12
  • If you have a huge number of predictors then a screening method can be useful to reduce the computational cost of the algorithm you are going to apply. Screening is just a synonim here for feature selection. It can also be useful to gain some understanding of you model. This is the main difference with PCA or similar dimensionality reduction methods: note that with PCA you proyect into a lower dimensionality space but the new features are meaningless there. Summing up, with feature selection you get relevant features (insight of your problem) plus less computational cost. – skd May 25 '15 at 16:02

0 Answers0