LoginSign Up
← Back to all sparks
F

fastml

ANALYTICS
Velocity0.0

R package fastml by selcukorkmaz — release notes from GitHub.

github.com/selcukorkmaz/fastml ↗

fastml added survival modelling and leakage-proof resampling, moving past classification and regression.

automltidymodelssurvival analysiscross-validationdata leakager package
◆Current state
A tidymodels-based AutoML wrapper that trains, tunes and compares many engines from one call. The 0.6.x line added engine-specific tuning parameters, class-imbalance handling, early stopping and DALEX-based explainability. The 0.7.5 release is far larger: a full survival analysis task with its own engines, MICE imputation and integrated Brier scoring, plus unbiased nested cross-validation, grouped, blocked and rolling resampling helpers, fold-wise imputation, recipe leakage checks, and a sandbox for user-supplied preprocessing.
◆Where it's heading
The package is moving from convenience wrapper to something that has to be defensible statistically. Nested cross-validation, fold-wise rather than up-front imputation, and explicit leakage checks are all corrections to the shortcuts that make AutoML easy and its scores optimistic. Survival adds a third task type alongside classification and regression, and it arrived with its own metrics rather than being bolted onto the existing ones. Note the entry body is cut off at 8,000 characters, so the release is larger than what is shown.
◆Prediction
Expect the remaining survival engines to fill in and the sandboxing of custom preprocessing to tighten, since both were still being iterated on within this same release's commit list.

◆Recent moves

  1. 8mo ago

    Version 0.7.5

    ⚡ SPARK

    Survival analysis becomes a first-class task with its own engines and integrated Brier scoring, and the resampling machinery is rebuilt around unbiased nested cross-validation, grouped and blocked schemes, fold-wise imputation and explicit leakage checks. The release that makes fastml's numbers defensible rather than just fast.

    View source ↗
  2. 1y ago

    Engine-specific tuning, imbalance handling and explainability

    Adds engine-specific tuning parameters and a no-tuning strategy, class-imbalance handling, early stopping, a verbose progress option, and multiple DALEX-based explainability methods. Broadens control over how models are fit rather than which tasks are possible.

    View source ↗
  3. 1y ago

    Single-workflow evaluation fix

    A stable CRAN release with a fix to single-workflow evaluation and a refactor of process_model() to take training data and a metric. Small corrections ahead of the larger 0.6.2.

    View source ↗
  4. 1y ago

    Version 0.5.0

    A CRAN release marker with no notes beyond the version. Nothing about its contents can be read from the feed.

    View source ↗