LoginSign Up
← Back to all sparks
M

mlr3extralearners

ANALYTICS
Velocity0.0

Extra learners for use in mlr3.

mlr3extralearners.mlr-org.com ↗

The mlr3 learner catalogue is growing fast and pruning hyperparameters just as deliberately.

mlr3learner-catalogh2ohyperparameterspython-interop
◆Current state
mlr3extralearners is the overflow catalogue for mlr3 learners that do not ship in the core packages — currently spanning H2O, Botorch, fastai, glmnet, survival and competing-risks models. The last two feature releases added roughly thirty learners between them. 1.6.0 then went the other way, cutting hyperparameters that were never correctly forwarded.
◆Where it's heading
Two forces are visible. The catalogue expands in bursts — 1.4.0 and 1.5.0 each added large batches, including a full H2O family and Bayesian regression models — while the maintenance releases in between are dominated by skipping tests on platforms where Python-backed learners crash. 1.6.0 marks a shift toward correctness of the existing surface: priority_lasso parameter sets reduced to what actually passes through, and Cox-inapplicable glmnet parameters removed.
◆Prediction
The Python-backed learners are the recurring source of platform instability, so expect continued pinning and test-skipping there alongside the next batch of additions.

◆Recent moves

  1. 1mo ago

    Hyperparameter sets pruned where arguments were never forwarded

    priority_lasso learners drop the glmnet pass-through parameters that were not consistently forwarded, and surv.glmnet and surv.cv_glmnet shed parameters inapplicable to the Cox family. Breaking, but in the direction of a parameter set that means what it says.

    View source ↗
  2. 3mo ago

    Dependency version updates

    Switches to the CRAN release of survdistr and to mlr3cmprsk 0.0.5, with crs parameter updates.

    View source ↗
  3. 4mo ago

    Platform-specific test skips

    A release composed entirely of skipping tests where Python backends crash or time out — fastai, botorch, tabpfn and h2o.glm across Windows and macOS.

    View source ↗
  4. 4mo ago

    Sixteen new learners, including a full H2O family

    Adds H2O GLM, GBM, random forest and deeplearning learners for both classification and regression, plus a competing-risks Coxboost, a Botorch fully Bayesian model and several neural network learners. Two survival learners move to the attic.

    View source ↗
  5. 6mo ago

    Twenty new learners and a survival learner rename

    A large batch spanning Bayesian treed models, partial least squares, regularised regression and neural networks, with surv.flexible renamed to surv.flexsurvspline to match the Royston/Parmar model it actually wraps.

    View source ↗