mlr3learners 0.5.4

mlr3learners 0.5.3

Survival learners have been moved to mlr3extralearners (maintained on Github): https://github.com/mlr-org/mlr3extralearners

Most learners now reorder the columns in the predict task according to the order of columns in the training task.
Removed workaround for old mlr3 versions.

eval_metric() is now explicitly set for xgboost learners to silence a deprecation warning.
Improved how the added hyperparameter mtry.ratio is converted to mtry to simplify tuning.
Multiple updates to hyperparameter sets.

Fixed the internal encoding of the positive class for classification learners based on glm and glmnet (#199). While predictions in previous versions were correct, the estimated coefficients had the wrong sign.
Reworked handling of lambda and s for glmnet learners (#197).
Learners based on glmnet now support to extract selected features (#200).
Learners based on kknn now raise an exception if k >= n (#191).
Learners based on ranger now come with the virtual hyperparameter mtry.ratio to set the hyperparameter mtry based on the proportion of features to use.
Multiple learners now support the extraction of the log-likelihood (via method $loglik()), allowing to calculate measures like AIC or BIC in mlr3 (#182).

Changed hyperparameters of all learners to make them run sequentially in their defaults. The new function set_threads() in mlr3 provides a generic way to set the respective hyperparameter to the desired number of parallel threads.
Added survival:aft objective to surv.xgboost
Removed hyperparameter predict.all from ranger learners (#172).

Fixed a potential label switch in classif.glmnet and classif.cv_glmnet with predict_type set to "prob" (#155).
Fixed learners from package glmnet to be more robust if the order of features has changed between train and predict.

The $model slot of the {kknn} learner now returns a list containing some information which is being used during the predict step. Before, the slot was empty because there is no training step for kknn.
Compact in-memory representation of R6 objects to save space when saving mlr3 objects via saveRDS(), serialize() etc.
glmnet learners: penalty.factor is a vector param, not a ParamDbl (#141)
glmnet: Add params mxitnr and epsnr from glmnet v4.0 update
Add learner surv.glmnet (#130)
Suggest package mlr3proba (#144)
Add learner surv.xgboost (#135)
Add learner surv.ranger (#134)

Split glmnet learner into cv_glmnet and glmnet (#99)
glmnet learners: Add predict.gamma and newoffset arg (#98)
We now test that all learners can be constructed without parameters.
A new custom “Paramtest” which lives inst/paramtest was added. This test checks against the arguments of the upstream train & predict functions and ensures that all parameters are implemented in the respective mlr3 learner (#96).
A lot missing parameters were added to learners. See #96 for a complete list.
Add parameter interaction_constraints to {xgboost} learners (#97).

Added learner classif.multinom from package nnet.
Learners regr.lm and classif.log_reg now ignore the global option "contrasts".
Add vignette additional-learners.Rmd listing all mlr3 custom learners
Move Learner*Glmnet to Learner*CVGlmnet and add Learner*Glmnet (without internal tuning) (#90)

Added parameter and parameter dependencies to regr.glmnet, regr.km, regr.ranger, regr.svm, regr.xgboost, classif.glmnet, classif.lda, classif.naivebayes, classif.qda, classif.ranger and classif.svm.
glmnet: Added relax parameter (v3.0)
xgboost: Updated parameters for v0.90.0.2

Fixed a bug in *.xgboost and *.svm which was triggered if columns were reordered between $train() and $predict().