mlr3pipelines

mlr3pipelines copied to clipboard

E.g. supersampling, smote etc. -- currently the best we do is throw an error? The more widely used things like weights should probably be supported.

mb706

Resampling: cv-plus

2

@berndbischl

sebffischer

Implement a `PipeOp` quantile prediction for learners that can only predict one quantile like gbm https://github.com/mlr-org/mlr3extralearners/blob/0297a064dda0424305a21f4b02cc27b47438f354/R/learner_gbm_regr_gbm.R#L105-L107. The pipop should train model for each requested quantile in `$train()` and then combine...

be-marc

```r 84/113 [05-pipelines-branch-03-prep] 85/113 86/113 [05-pipelines-branch-03] Error: ! Cannot infer active output of PipeOpBranch brnchPO with non-string 'selection'. ```

sebffischer

Fixes `PipeOpImputeOOR` handling NAs only during predict

2

closes https://github.com/mlr-org/mlr3pipelines/issues/894

advieser

POs that handle supervised tasks: do they handle derived tasks (RegrST etc) well?

3

we should check, maybe build a test case. also: give useful error message when POs can handle only classif & regr but nothing else

mb706

adas from themis

1

Probably less error prone than smotefamily, see #885

mb706

check_types should give more user-friendly error messages

1

```r preproc(iris, po("smote")) ``` should make it obvious what the problem was: input to graph incompatible.

mb706

Autotest that setting paramset to `$default` does not change its behaviour.

mb706

It does not, because it does not inherit from PipeOpTaskPreproc

sebffischer