Hello colleagues,
I have seen researchers in my area (biometrics) mostly use the following feature selection methods:
Info gain, cfs subset selection, mutual information, KS test, or PCA or any other state of the art ?
Which should be used when ? I generally support the selected features by applying more than one methods and see if both of them are selecting almost the same features. How correct this argument is ?
Thanks in advance!Ask
Thanks Michel. The book seems interesting!
I found a paper http://www.cs.waikato.ac.nz/~mhall/HallHolmesTKDE.pdf
which states the same thing.
PCA and LDA are both fairly common methods for discriminating between different features to show which are the most relevant.
Dear Blair, thanks for the response. How can you compare PCA, LDA or any other features selection and transformation methods I have mentioned ? Is there any source you can point out which talks about which method is better in which scenarios.
Hi Rajesh, If you check out the No Free Lunch Theorem in AI you will see that different kinds of classifiers will be most suitable in different domains. Something similar may be going on here. Part of your research contribution may be to produce a suitable dataset with ground truth information and run an experiment to see which works best with your data. This should be publishable and interesting to others working in the field (i.e. get you some citations).
Dear James,
Thank you for your interesting reply. I will definitely look into that theorem.
- Badges
- Science topic
- Similar topics
- Biological Science
- Bioecology
- Systems Ecology
- Landscape Ecology
- Connectivity