**🎉 🎉 RAPIDMINER 9.10 IS OUT!!! 🎉🎉**

### Download the latest version helping analytics teams accelerate time-to-value for streaming and IIOT use cases.

## CLICK HERE TO DOWNLOAD

# Heuristic for choice of parameter C in LibSVM

Hi there,

I was trying to find more info on the heuristic choice of the parameter C used by RapidMiner (i.e. the one used when the parameter is given as 0). However, the only place I could find mentioning this heuristic was this old thread from this forum. Is there any other place where this heuristic is discussed? I could not find the references cited by Ingo on that thread. Hastie and Tibshirani have far too many publications to search them one by one. Perhaps someone could give me any directions for justifying this heuristic value?

Cheers,

Dann

I was trying to find more info on the heuristic choice of the parameter C used by RapidMiner (i.e. the one used when the parameter is given as 0). However, the only place I could find mentioning this heuristic was this old thread from this forum. Is there any other place where this heuristic is discussed? I could not find the references cited by Ingo on that thread. Hastie and Tibshirani have far too many publications to search them one by one. Perhaps someone could give me any directions for justifying this heuristic value?

Cheers,

Dann

0

## Answers

21MavenCherkassky: Practical Selection of SVM Parameters and Noise Estimation for SVM Regression

I would be happy if you let me know if it worked...

Milan

5Contributor IICheers,

Dann

849Guru5Contributor IIThanks for replying, but what exactly do you mean by posting a link to the guide by Hsu, Chang and Lin? I could have overlooked it, but I didn't find any reference about the heuristic used by RapidMiner on it. Would you care to explicitly quote where they mention the heuristic described by Ingo Mierswa on this thread?

Regards,

Dann.

849Gurutheirguide to setting parameters fortheir librarymight be relevant.5Contributor III am aware that the LibSVM library is written by Chang and Lin. But actually, the heuristic parameter seems to be set by RapidMiner, not by LibSVM. At least this was the impression I had by reading Ingo's answer. Sorry for not making this clear. Thanks for the reply, by the way.

Regards,

Dann

849GuruReading that thread myself I found it unclear, but figured that whetever RapidMiner does it must depend on the properties of the underlying functions. If you search this forum you'll see that even the meaning of 'C' in SVMs gets vague, so zero 'C' ....

That's why I like Association Rules for booleans!

5Contributor IIThanks for the observation. In either case, I am starting to thing that this heuristic arises from the Wahba, Lin, and Zhang bounds on the leave-one-out cross-validation error. A paper by Olivier Chappele [1] states that the Wahba, Lin, and Zhang bound is such that T = sum( a_i k'(x_i, x_i ) ), in which k' is assumed to be an augmented kernel function with the parameter C embedded within it so that k'(x_i, x_i) = k(x_i, x_i) + 1/c. All sums are from 1 to n, in which n is the number of samples in the data set.

Now, if we go on and substitute k' with the aforementioned relation, I guess we can arrive on

T = sum( a_i k(x_i, x_i) ) + sum(a_i 1/c).

Since we would like to find the minimum of

T, I suppose a straightforward strategy would be to assume thatTis dependent on the Lagrange multipliersa_i, derive in respect toaand impose stationarity. This would lead to:del T / del a = sum( k(x_i, x_i) ) + sum (1/c) = sum( k(x_i, x_i) ) + n/c = 0

And with a little algebra this can result in expression for C very similar to the heuristic suggestion as C = n / sum(K(x_i, x_i), albeit with the sign for C inverted. Something else may still be missing.

Cheers,

Dann

[1] http://research.microsoft.com/en-us/um/people/manik/projects/trade-off/papers/ChapelleML02.pdf