Using cross validation with filter or embedded method produce in each fold different rank list depending on the number of sample. So what is the final ranking list if I use svm-rfe to rank my features with 10 fold.
Someone told me we can take the average rank, or highest frequency features from each fold.
Hello Dheeb,
If I'm not mistaking, standard ranking by Weka does exactly the thing that "someone told you", i.e. averages feature rankings over all folds.
You could play with several metrics, for instance, you could produce final ranking based on a maximum-per-fold rank for each feature.
There should be average rank of all folds for cross validation accuracy.
OR
You can use median of all folds as a final cross validation accuracy.
hello friend
In 10 fold CV , the final rank is the average of all the 10 folds.
Its not the median......
I have done some works related to this concept.
see my papers for more details.
All the best.
From,
Dr.Indrajit Mandal
Thank you very much for your suggestions,
but, if we want to know the name of these features some time it become uneasy to know which features are the most important features even if we run the experiment 10 times.
- Badges
- Science topic