JSVMLearner with non linear kernel

ilaria_gori

Could you explain me why in jSVMLearner with non linear kernel, the vector w is calculated as linear combination of the training vectors, as in case of linear kernel? Why it is not simply given the function value which is sufficient to have the output?
thanks!

ilaria

haddock

Has this not been dealt with before?

http://rapid-i.com/rapidforum/index.php/topic,1455.msg5535.html#msg5535

ilaria_gori

thanks, but it's a different problem,

ilaria

land

Hi,
I don't think haddock referred to the question in this topic. I rather believe he points to the solution I recommended: Reading the original paper.

Greetings,
  Sebastian

ilaria_gori

thanks, but I already read the paper and in fact I learnt that the vector w should be calculated in another way, that is with phi(x_i) and not with the x_i, as Rapid Miner does.

ilaria

land

Hi,
that's correct. So where did you see, that w was calculated in the wrong way? Did you see it in the source code or was it displayed within rapid miner?

Greetings,
  Sebastian

ilaria_gori

Hi,
I simply calculated it by myself, in order to see if I had understood the algorithm, and I remarked that Rapid Miner calculates it with the x_i and not with the phi(x_i).  It's not a problem if you use only the confidence as output, but I thought It was interesting to point it out,

ilaria

haddock

G'Day Folks,

Not sure Ilaria got due credit here for the close work, so let me be the first to commend it to you, top stuff!

land

Hi again,
where exactly is the w shown in RapidMiner? I just want to take a closer look. If I remember correctly, the phi cannot be expressed for any kernel?

Greetings,
  Sebastian

ilaria_gori

Hi, you are right, the phi cannot be expressed for all kernels. Probably RM could simply not to give the w, which now is given in the training process output model, because it's not necessary.
greetings,

ilaria