Hello,
I have a problem in feature selection I would be thankful if you can help
me.
I have a dataset with limited samples (for example 100) and a lot of
features (for example 3000) and i have to do feature selection.
if i use cross validation (for example *10 fold*) i rank the features based
on 90 samples (using svmrfe method) i achieve ranked feature for example
{f2,f4,f1,f3,...} (it means f2 is ranked first with svmrfe) now, I want to
know how many features i should use? so i should compute the performance for
n feature selected from first of the ranked list and compute the performance
of it. for example train learner with f2, another time with f2,f4, another
time with f2,f4,f1 ... and see which is better , but my problem is:
1) first of all,for comparison should i use the performance of 9 fold that
has been used for ranking or the performance of learner on the fold which
has been left out?I mean in the *feature selection* *step (not in the final
evaluation)*,for example to see I should select only f2 or select {f2 , f4}
how should I compare?
2) in each stage of cross validation different feature subset will be
created . i can compute for each feature the number of times it has repeated
in each folding result, but after that how can i conclude the final feature
set?
can you please help me? I need your urgent help.
thanks in advance
Azadeh
[[alternative HTML version deleted]]
