I have a combination of features extracted from 3 descriptors, namely GLCM based feaures(correlation, homogeneity,energy and contrast ), Local binary patterns (256) and discrete wavelet transform coeficcnets. and a dataset of 40 cases. and I'm using SVM with RBF kernel.

would there be a need for the feature selection? if so, what do you recommend


I'd recommend assessing the performance of the model "as-is" and then decide for feature selection or not.

If you have good performance for both train / test sets, then there's no need to perform feature selection.

If you end up overfitting your data (low train error, high test error) you may want to reduce the number of features using feature selection / regularization (the cost function is penalized when the model uses a large number of parameters)

If you have low train performance, this means that you have high bias (the model does not learn properly from those features). This usually happens if the model is not complex enough to extract structure from the data. You may want to add features in this case, rather than selecting them out.

Hope this helps.

  • $\begingroup$ Thank you so much sir. will do as you suggested. If may i ask, do you think SVM with RBF kernel is suitable or simple linear SVM. My dataset is 40 cases which has binary class. $\endgroup$ – gin Dec 6 '18 at 13:45
  • $\begingroup$ I'd suggest starting with the simple SVM and then trying out different kernels or even other algorithms. You may even start with a logistic regression at first. Do not forget to scale your features in this case. If the answer was helpful, fell free to upvote it ;) $\endgroup$ – seiseman Dec 6 '18 at 14:03

Your Answer

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

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