I will include your bug fix into the code.

Thanks again and best regards,
mIHA

Thanks for a wonderful wrapper for SVMLight. The API is very useful and easy to work with. The featureVector unloading routine has a memory leak (detected by VLD). In “void DeleteFeatureVector(int id)” call, you forgot to delete the feature_vector itself. Just adding “delete feature_vector;” before calling “feature_vectors.erase” did the job.

Hope this helps
cheers
Vivek

If you have no knowledge of C# and you plan to use SvmLightLib on Linux, I suggest you simply use C++ (alternatively, you could try C# under Mono but I guess this would only complicate things). The core of SvmLightLib is implemented in C/C++ anyhow, so it should compile with gcc (probably with minor modifications). I don’t have the time right now to prepare a C++ usage example. But I will do it as soon as the time permits.

Let me know if you manage to compile it with gcc : )

Best regards,
mIHA

I guess you get this error message when you try to reference svmlightlib.dll from the project. You shouldn’t explicitly reference it at all. You simply put the file in a folder on your disk (e.g. c:\svmlightlib\bin) and add the folder to the environmental variable PATH. To do this, you need to go to Control Panel/System/Advanced/Environmental variables and look up PATH under System variables. Double-click the entry and append something like “;c:\svmlightlib\bin” to the variable value. You might need to restart your computer; you will definitely need to restart the IDE.

Best regards,
mIHA

I wonder how to run the demo (in c#) in Linux. I have no knowledge about C# at all.

Also, I’d like to know how to make use of your library in C/C++ (in Linux environment).
Could you give an example?

Thanks a lot!!

There are at least two good reasons why you should optimize your model (i.e. do your feature selection) on *the validation set* (i.e. part of the training set that you don’t use for training):

1.) To avoid overfitting. Overfitting means that the model performs well on the training set, but (much) worse on the test set. So, you need to select those features that increase the model’s performance on the validation set therefore avoiding overfitting to the training set.
2.) The model is useless if it doesn’t perform well on the test set : ) You can get 100% accuracy on the training set and it doesn’t mean anything if the accuracy is below 50% on the test set. If you optimize your model’s performance on the validation set, you make sure that your model will perform well in practice.

And I have another doubt about your idea on margin maximization. In my opinion, if you build one model with one feature set (model A) and then another one with another feature set (model B), the size of the margin is not directly comparable. It can happen that the model A performs better in practice despite the fact that its margin is smaller than that of the model B. I’m not 100% sure about this, but you should definitely check in the appropriate literature.

Best regards,
mIHA

so I feed the new descriptors to an svm and the idea is to find the “best” b vector that gives me the biggest margin (equivalently the lowest w^{T}*w)

I wanted the c# version since my code is in emgucv. Anyway, I save everything in a file and I am developing a java version with libsvm-java which i “tweaked” to provide me the objective function. PSO is indeed particle swarm optimization.

Maybe the I am wrong but the measure accuracy is on test data so it is not an option to optimize -> overfitting, the margin is better.

Regards
Vasileios

PSO, I guess, stands for the particle swarm optimization method? So, the objective function should reflect the quality of the model. One possibility is to perform evaluation (e.g. in a 10-fold cross validation setting) and measure accuracy, F1, or any other standard evaluation metric as your objective function for PSO.

Let me know if I misunderstood your inquiry.

Best regards,
mIHA