Huller is a simple and efficient online SVM approximation method published by Antoine Bordes and Leon Bottou. This page provides a simple reimplementation in ANSI C of their original paper. I found their intuitive approach quite elegant, understandable and easily implementable.

If you are looking for a state of the art SVM implementation, you most likely want to use the excellent libsvm library instead, which is a soft margin SVM implementation.


All the documentation that is available is included in the huller.h interface file. Additionally, I recommend the original Huller paper. An example is provided with the sources.

But here is a beautiful shot of the output, plotted with gnuplot. It shows a simple toy 2D spiral example, separable by a Gaussian kernel (sigma = 0.5).
Spiral dataset, decision function

Note the blue zero contour plotted on the ground surface, which is the decision boundary. The yellow and pink level sets approximately contain the negative and positive samples, respectively.


The latest available version is huller-0.1.tar.gz, added 2006/01/22. It is licensed under the very liberal MIT license, so you can do almost anything with it.


For questions, suggestions and constructive criticism you can reach me at nowozin@gmail.com.