Hi all, I hope you don't find this question boring or irrelevant. I am very interested in SVM. I realize SVM research mainly consists of two parts - the formulation (which is concerned with the choice of regularization and loss function) and the algorithm to solve the formulation (e.g. gradient descent, coordinate descent, trust region newton method). I have also done a quick survey on available SVM methodologies/packages out in the public: - Liblinear runs very quickly in sacrifice of a little bit of accuracy. - Ramp loss LP SVM uses a relatively innovative loss function known as "ramp loss function" and applies DC programming (difference of convex functions) to solve the formulation, among other techniques. - L1-npsvm (non-parallel proximal support vector machine) uses two hyperplanes instead of one (which is the traditional setting) to achieve amazingly high classification accuracy. Do you have any comment/opinion on the above? Also, do you know any active research direction on SVM? Please share with me if any. Have a good day! Sincerely, Mr. F
asked
bookyeah1679 |

Thank you for your reply Petter. I tried libsvm before, and it seems that liblinear is a continuance of libsvm by National Taiwan University. Both can achieve similar classification rate, but liblinear runs much quicker.
answered
bookyeah1679 |