Thanks for tackling this, @alexshacked ! Wouldn't enforcing that all sample_weights are positive on the Python side (and raising an error if they are not) be a simpler solution?

Thanks for tackling this, @alexshacked ! Wouldn't enforcing that all sample_weights are positive on the Python side (and raising an error if they are not) be a simpler solution?

Yes, @rth. enforcing that all sample_weights are positive on the python side, in BaseLibSVM.fit() for example would be simpler. This was my original intention.
Until I saw that in the algorithm implementation in svm.cpp:2342, the negative weights are actually handled.
PREFIX(model) *PREFIX(train)(const PREFIX(problem) *prob, const svm_parameter *param, int *status) { PREFIX(problem) newprob; remove_zero_weight(&newprob, prob);
This is the current implementation in the code base. There was a concious decision by the SVM implementor to handle input with negative/zero weights, by removing this samples from the input, and training the model on the remaining samples. The SVM implementor was not shy about rejecting invalid input and in function check_parameter() it does just that for many scenarios. But in the case of negative weights the decision was not to reject the input but to handle it by removing samples with negative weights.
The only thing is that while doing so the SVM implementor ommited some border cases that arise, like the one discovered by bug #9494.
My thought was that a less intrusive fix would be to teach the implementation to handle the border cases without changing current behavior of the SVM algorithm which handles input with negative weights.
Hence, the present solution.
If on the other hand, the decision for SVM algorithms would be not to handle input where part of the samples are zero/negative anymore, I would be happy to implement. On the python side naturally:-)

Hi @glemaitre. I don't know what self-request a review means. But, I"ll be glad if you will review my PR.
I see that you also have worked with sample weights lately
If you need me to do something to enable you to review please tell me.

self-request a review means

@alexshacked it is just for me to track the PR which I am going to review :)