classif.plsdaCaret sometimes does not terminate

[mb706]

... because it uses the "oscorespls" algorithm, which is prone to numerical instability. The pls package paper mentions that the "kernelpls" algorithm gives the same result and is usually faster. I suggest changing it here.

library(fastICA)
> dm <- fastICA(matrix(runif(6000), ncol=6), 6, method="C")$S
> mkt <- makeClassifTask("orthotask", data.frame(t=sample(factor(c('a', 'b')), 1000, replace=TRUE), dm), target='t')
> train(makeLearner('classif.plsdaCaret'), mkt)  # runs forever

[berndbischl]

well, i kinda see your point, but i really would not like to change the defaults of every learner in mlr.
(in a manual way)

does this break always or just sometimes?

[berndbischl]

ok, expand a little bit on your argument, then possibly do a PR with a cleanly documented "note" in the learner why this is changed in mlr

[mb706]

"kernelpls" is the default behaviour (see 'plsralg' in pls.options), the method = "oscorepls" in trainLearner.classif.plsdaCaret overrides this.

[berndbischl]

ok, so someone HARDCODED a method in the mlr learner, which is NOT even the DEFAULT in the underlying package?

caret::plsda(d$data, d$target, method = "oscorespls", ...)

then by any means please change and correct this!

also there is no "method" apparrantly oin parset