SVC(kernel='linear') supports sample weights just in case ...

Yes, but I am working on sparse data and I wish to use the decision_function method. Since that method is not implemented for svm.sparse.SVC..i have to use LinearSVC...is there any other way to overcome this?

LinearSVC does not rely on libsvm, but on liblinear. Thus it would be necessary to add support for sample weights to liblinear, which is a sizeable amount of work.

Well, according to the following link
http://www.csie.ntu.edu.tw/~cjlin/libsvmtools/#weights_for_data_instances
liblinear also has that feature .. can that be somehow used?

No, that's the LibSVM docs. To repeat what Gael said, LinearSVC is built on top of LibLinear, by the same authors.

Okay. Thanks.

There does seem to be a LIBLINEAR fork supporting sample weights... http://www.csie.ntu.edu.tw/~cjlin/libsvmtools/#weights_for_data_instances

Supporting sample_weight in LogisticRegression may also be helpful.

I just noticed the same comment was made by @aabbi, but I don't understand why it was rejected. The code there is indeed a modified version of LibLinear.

Seeing as scikit-learn has adopted the nonstandard LibSVM with sample weights, I assume the changes to LIBLINEAR should indeed be merged, so I'm reopening this Issue (and am willing to be shot down for doing so!). The differences can be viewed at http://github.com/jnothman/scikit-learn/tree/liblinear-sample_weight (particularly jnothman@54ec588). I have based this on the commit before Lars imported liblinear 0.91; there was no record of the diff from the downloaded LIBLINEAR 0.91 to that commit, which should have involved a branch and merge. I haven't (yet) modified our set_problem etc. nor merged in the later changes to linear.cpp, and am more than happy for someone else to do so :P

We already have class weights. You want to add sample weights in extra,
right?

Should the class_weights be then refactored to use the sample_weights? It
would be possible and might make the code simpler.

The trick business here will be to make really sure that all the patches
applied to liblinear are kept.

You want to add sample weights in extra, right?

For symmetry with LibSVM if nothing else.

Should the class_weights be then refactored to use the sample_weights?

It looks like that liblinear fork supports both independently.

The trick business here will be to make really sure that all the patches applied to liblinear are kept.

I've tried the merge that should apply them; there are 15 blocks with conflicts in linear.cpp and that's it. "Making really sure" hopefully is a matter of tests passing :s

fixed at some point since then.