R package mlr: option for custom kernels in classif.ksvm

Question

I want to build an svm with a custom kernel. Usually I use the R package kernlab for that. Since, I want to try out different kernels and tune the hyper-parameters, I wanted to use the nice package mlr. However, as far as I can see, it doesn't support the kernel type option "matrix" to pass a custom kernel to the ksvm learner ("classif.ksvm").

Does somebody know if there is a plan to fix that? Or if there is another package which allows custom kernels and nice wrappers for tuning parameters and resampling methods. To my knowledge, the caret package is also not taking custom kernels.

For similar cases mlr has `configureMlr(on.par.out.of.bounds = "quiet")`. But in this case the argument `kernel = "matrix"`changes the behavior in a way that you have to change the learner like @Lars Kotthoff did. — jakob-r, Jan 16 '17 at 09:27

score 3 · Accepted Answer · answered Jan 13 '17 at 18:00

We don't have any plans to support that. You can define a custom learner to support this quite easily though. There are two changes to classif.ksvm as far as I can see (untested).

First, allow the new parameter value for the kernel parameter:

makeDiscreteLearnerParam(id = "kernel", default = "rbfdot",
    values = c("vanilladot", "polydot", "rbfdot", "tanhdot", "laplacedot", "besseldot", "anovadot", "splinedot", "matrix"))

Then, change the train function to take the new kernel into account:

trainLearner.classif.ksvm = function(.learner, .task, .subset, .weights = NULL, degree, offset, scale, sigma, order, length, lambda, normalized,  ...) {
  kpar = learnerArgsToControl(list, degree, offset, scale, sigma, order, length, lambda, normalized)
  f = getTaskFormula(.task)
  pm = .learner$predict.type == "prob"
  parlist = list(...)
  if (base::length(kpar) > 0L)
    kernlab::ksvm(f, data = getTaskData(.task, .subset), kpar = kpar, prob.model = pm, ...)
  else if (parlist$kernel == "matrix")
    kernlab::ksvm(kernlab::as.kernelMatrix(getTaskData(.task, .subset)), data = getTaskData(.task, .subset), prob.model = pm, ...)
  else
    kernlab::ksvm(f, data = getTaskData(.task, .subset), prob.model = pm, ...)
}

This is assuming that the data you're passing in the task defines the custom kernel, which is a bit kludgy...

can you kindly give a tutorial on how to invoke jaccard kernel in svm? — KHAN irfan, Aug 29 '17 at 19:08

R package mlr: option for custom kernels in classif.ksvm

1 Answers1