SVM Practical (How to get good results without cheating)
Jason Weston, Arthur Gretton and Andre Elisseeff
In this practical we will look at how to get good experimental results (that don't involve cheating!) with support vector machines. We will cover a few obvious things for people new to
machine learning / SVMs which aren't always covered in typical publications/ talks. We will cover
three main aspects:
Understanding SVMs to get good results (tricks of the trade).
Special kernels for special problems (encoding prior knowledge.)
Proper experimental design (to avoid cheating/ biased results.)
There are a number of pieces of code in Matlab that serve as exercises. They
use the Spider Machine Learning Library to train the SVMs and other systems, the
Spider is freely available at: http://www.kyb.tuebingen.mpg.de/bs/people/spider.
Slides (same as the original booklet, but includes solutions to the exercises)