LoginSign Up
← Back to all sparks
M

mlr3learners

ANALYTICS
Velocity0.0

Essential machine learning algorithms wrapped as learners for mlr3

github.com/mlr-org/mlr3learners ↗

mlr3learners spends its releases absorbing upstream churn

mlr3machine-learningr-statslearnerscompatibilityxgboost
◆Current state
mlr3learners wraps the standard model packages — ranger, xgboost, glmnet, kknn — for mlr3. Its recent history is dominated by upstream events rather than its own plans: kknn was pulled from CRAN and its learners removed in 0.11.0, then restored in 0.12.0 when the package returned. The newest release absorbs glmnet 5.0 while adding a predict_raw flag across all learners and probit support to logistic regression.
◆Where it's heading
The package's job is insulation, and the changelog shows what that costs — compatibility-only releases interleaved with small capability additions that expose more of each upstream model. The direction of travel is toward giving users access to the raw upstream objects rather than hiding them.
◆Prediction
Expect the next release to track another upstream version bump, with incremental exposure of learner-specific fields continuing alongside.

◆Recent moves

  1. 2mo ago

    predict_raw across all learners, plus probit and ranger internals

    The broadest release in this window: every learner can retain raw upstream prediction objects, logistic regression gains a link parameter for probit, and ranger learners expose their native model and local importance. Continues the trend of exposing rather than hiding the wrapped models.

    View source ↗
  2. 8mo ago

    xgboost 3.1.2.1 compatibility

    A compatibility-only release with a single line of change. Representative of how much of this package's cadence is upstream-driven.

    View source ↗
  3. 9mo ago

    Uncertainty estimation methods for ranger regression

    Adds ensemble standard deviation and law-of-total-variance uncertainty estimates to regr.ranger, and lets ranger's out-of-bag error be stored without keeping the model. Real predictive capability rather than compatibility work.

    View source ↗
  4. 10mo ago

    Development snapshot: LDA test adjustment

    A development-version tag containing a single test change. Not a user-facing release.

    View source ↗
  5. 1y ago

    kknn learners restored after returning to CRAN

    Re-adds classif.kknn and regr.kknn, reversing the removal forced one release earlier. Read alongside 0.11.0 this is a restoration, not a new capability.

    View source ↗
  6. 1y ago

    kknn learners removed after CRAN archival

    A breaking removal forced by kknn's departure from CRAN, alongside mlr3 1.0.0 compatibility. The clearest illustration of how exposed this package is to upstream decisions.

    View source ↗