You don't really expect SVM to give you good performance with no parameter
tuning at all, do you?
Try:
m2 <- best.svm(class~., data=spiral, gamma=2^(-3:3), cost=2^(0:5))
plot(m2, spiral)
Andy
From: Joshua Gilbert>
> I'm using an SVM as I've seen a paper that reported extremely good
> results. I'm not having such luck. I'm also interested in ideas for
> other approaches to the problem that can also be applied to general
> problems (no assuming that we're looking for spirals).
>
> Here is my code:
> library(mlbench)
> library(e1071)
> raw <- mlbench.spirals(194, 2)
> spiral <- data.frame(class=as.factor(raw$classes),
> xx=raw$x[,1], y=raw$x[,2])
> m <- svm(class~., data=spiral)
> plot(m, spiral)
>
> You'll note that I have two spirals with 97 points each and I'm
using
> a kernel with a radial basis: exp(-gamma*|u-v|^2).
>
> You should be able to see a PNG of the resulting plot here:
> http://www.flickr.com/photos/60118409 at N00/91835679/
>
> The problem is that that's not good enough. I want a better fit. I
> think I can get one, I just don't know how.
>
> There's a paper on Proximal SVMs that claims a better result. To the
> best of my knowledge, PSVMs should not outperform SVMs, they are
> merely faster to compute. You can find the paper (with the picture of
> their SVM) on citeseer:
> http://citeseer.ifi.unizh.ch/cachedpage/515368/5
> @misc{ fung-proximal,
> author = "G. Fung and O. Mangasarian",
> title = "Proximal support vector machine classifiers",
> text = "G. Fung and O. Mangasarian. Proximal support vector machine
> classifiers.
> In F. P. D. Lee and R. Srikant, editors, KDD",
> url = "citeseer.ifi.unizh.ch/515368.html" }
>
> I don't have much of a background in SVMs, I'm learning as I go, so
> please don't hold back 'simple-minded' suggestions.
>
> I'm also asking the authors, but I'm not expecting a reply from
them.
>
> There was a paper by Lang and Whitbrock in 1988 (Learning to Tell Two
> Spirals Apart) that solved the problem with a neural network, but they
> used a very specialized network architecture. I would say that
> discovering such an architecture and then optimizing it would be very
> time-intensive.
>
> Thank you for any response.
>
> Josh.
>
> ______________________________________________
> R-help at stat.math.ethz.ch mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide!
> http://www.R-project.org/posting-guide.html
>
>