|
Dear metaoptimize community, I would like to know which is the easiest way to implement a linear SVM solver, when replacing the regularization term I have not found an SVM implementation that makes it easy (i.e. documented feature) to replace the regularization term. Which implementation/hack would you suggest ? ps: If you wonder why I would need this, you can for example consult equation 17 of this paper |
|
If your matrix in the regularizer is positive definite you can take its square root ( http://en.wikipedia.org/wiki/Square_root_of_a_matrix ) and fold it into the feature vectors of the examples, whih should allow you to reduce solving an SVM problem with an arbitrary norm to solving a standard SVM problem. That did the trick indeed, using the pseudo-inverse of the square root to transform the data, and then transform the learned weights back into the original data domain. Thanks for the suggestion !
(Feb 19 '14 at 16:51)
Rodrigo Benenson
|
