S3 internal method: getBaggedPredictions / generalize se estimation for bagged learners

[jakob-r]

For all learners that use internal bagging like ranger, randomForest, extraTrees, xgboost(?) we should have at least an internal S3 function like that: (really simple for randomForest)

getBaggedPredictions.regr.randomForest.model = function(.model, .newdata) {
  m = getLearnerModel(.model)
  predict(m, newdata = .newdata, predict.all = TRUE)
}

It returns a matrix with nrows = number of observations and ncol = number of bagged predictors. In predictLearner it can than be used if .learner$predict.type == "se" to call a se method like jackknife or simple standard error (more complex things like bootstrapStandardError need further thought):

getSePrediction(.learner, .model, .newdata) will be a function that simply switches depending on .learner$par.vals$se.method to call a generic function like getSeJackknifePrediction(.learner, .model, .newdata) which can simply rely on getBaggedPredctions(.model, .newdata) to calculate it's jackknife se estimation.

Side remark: Theoretically you could solve it without the S3 getBaggedPredictions and just generate the desired matrix within the code of predictLearner.regr.randomForest but I think it is handy to have this kind of separation if you want to do some fancy stuff later.

(I am quite certain that there have to be related issues)

[pat-s]

I do not see us implementing this in mlr at this stage. Good case for mlr3.