Does SVM perform poorly when fat-tailed data with outliers is used? What are some things that could be done to improve learning with such data? Does the choice of kernel and/or kernel parameter change?
For dealing with outliers, I was thinking of transforming the data such that values +/- 2 standard deviations are assumed to be outliers and would be replaced with the closest marginal values. (then transform the data into the range [-0.9, 0.9])
Assuming fat-tails are undesireable with SVM, what are some transformations that could be considered to help alleviate the fat-tails?
(I'm using LIBSVM with the Matlab interface)