Some Comments About Ch. 9 of Text
Page 371 of the text mentions that the maximal margin classifier of Sec. 9.1 can overfit when p is large. But this is also true of the
support vector classifier of Sec. 9.2, and the support vector machine of Sec. 9.3. If there are noise variables as predictors, then they will influence the
decision boundary ... there is no variable selection done with any of the Ch. 9 methods. Even though the noise variables ought to only have a mild influence on the
decision boundary, if the noise variables were eliminated before determining the
decision boundary (perhaps based on clues you obtained from trying other classification methods), then the boundary obtained might do a better job.
Ch. 9 omits a lot of the technical details (and with good reason ... they are very messy, and wouldn't be very meaningful for most readers of the book).
But if you're sufficiently interested, you might look into a couple of books by Vapnik. (Our text, ISL, doesn't give the references, but you can
find them in the back-of-the-book refrences in ESL.)