Least Square Support Vector Machine in Caret fail

Question

I try to fit a Least Square Support Vector Machine vie the caret package in R, but I am unable to get it to work. Even for extreme simple examples like this it fails:

library(caret)
library(tidyverse)
data("iris")


#to make this example a binary classification task

iris <- iris %>% filter(Species %in% c("setosa", "versicolor")) %>%
    mutate(Species = droplevels(Species))

svmls <- train(Species ~ .,
               iris,
               method = "lssvmLinear",
               preProc = c("center", "scale")
               )

with a couple of warnings like this:

In eval(xpr, envir = envir) :
  model fit failed for Resample09: tau=0.0625 Error in if (truegain[k] < tol) break : 
  missing value where TRUE/FALSE needed

While calling the lssmv function from kernlab directly succeed:

library(kernlab)
svmls2 <- lssvm(Species~.,data=iris)
svmls2

I would really appreciate any guess on what might be wrong.

Answer 1

I knew the question already old, but here is some answer

I also got the same error, when Look depth in it, The Caret Default of LSSVM Linear is using Polygonal Kernel, that looks like this:

getModelInfo()$lssvmLinear$fit
function(x, y, wts, param, lev, last, classProbs, ...) {
                    kernlab::lssvm(x = as.matrix(x), y = y,
                                   tau = param$tau,
                                   kernel = kernlab::polydot(degree = 1,
                                                             scale = 1,
                                                             offset = 1), ...)    
                  }

hence I edited it to use only the default kernel, this way it can behave like what expected:

newlssvm <- getModelInfo()$lssvmLinear
newlssvm$fit <- function(x, y, wts, param, lev, last, classProbs, ...) {
  kernlab::lssvm(x = as.matrix(x), y = y,
                 tau = param$tau)    
}

svmls <- train(Species ~ .,
               iris,
               method = newlssvm,
               preProc = c("center", "scale")
               )

I reclaimed that this issue was at kernlab, since this:

lssvm(Species~.,data= iris, kernel = kernlab::polydot(degree = 2,
                                                      scale = 0.01, offset=1))

give similiar errors like this:

Error in if (truegain[k] < tol) break : 
  missing value where TRUE/FALSE needed
In addition: Warning message:
In sqrt(G[kadv, kadv]) : NaNs produced