We recently saw "Why use SVM" hit the front page of HN, but I feel this is a much more rigorous treatment of the same topic.
In particular, Andreas does a great job discussing the relevance of the kernel to SVMs, its computational complexity, and methods to minimize that complexity via inner product space approximation.
Interesting to see how the Nystrom and Fourier approximations to the inner product space almost match the full kernel's accuracy in 1/3rd and 1/5th as much time, respectively. It is rare that one needs that extra bit of accuracy, so I imagine Nystrom would be far more appealing for anyone using sklearn.SVM.
In particular, Andreas does a great job discussing the relevance of the kernel to SVMs, its computational complexity, and methods to minimize that complexity via inner product space approximation.
Interesting to see how the Nystrom and Fourier approximations to the inner product space almost match the full kernel's accuracy in 1/3rd and 1/5th as much time, respectively. It is rare that one needs that extra bit of accuracy, so I imagine Nystrom would be far more appealing for anyone using sklearn.SVM.