LoginSign Up
← Back to all sparks
I

imbalanced-learn

AI-ASSISTANTS
Velocity0.0

Resampling techniques for imbalanced datasets

github.com/scikit-learn-contrib/imbalanced-learn ↗

The resampling companion to scikit-learn now ships mostly to stay compatible with it.

imbalanced-dataresamplingscikit-learncompatibilitypython
◆Current state
imbalanced-learn is at 0.14.2. Four of the six releases in the window exist to track a scikit-learn version — 1.5, 1.7, 1.8 and 1.9 in turn — or NumPy 2.0. The genuine additions are thin: InstanceHardnessCV in 0.14.0 and a clearer SMOTENC error when the categorical encoder collapses categories.
◆Where it's heading
The project has settled into the role of a compatibility shim with a stable sampler catalogue. Release timing is set by upstream scikit-learn, not by its own roadmap, and the deprecations queued in 0.13.0 show the surface narrowing rather than growing.
◆Prediction
The pattern points to the next release being another scikit-learn compatibility bump, with the Pipeline check_is_fitted deprecation scheduled to become an error in 0.15.

◆Recent moves

  1. 2mo ago

    scikit-learn 1.9 compatibility and a SMOTENC error message

    scikit-learn 1.9 compatibility plus one better error message in SMOTENC when the categorical encoder drops or merges columns. No change to the sampler catalogue.

    View source ↗
  2. 7mo ago

    scikit-learn 1.8 compatibility release

    A single-line release tracking scikit-learn 1.8. It exists so installs do not break, and nothing else.

    View source ↗
  3. 0y ago

    InstanceHardnessCV splits folds by sample hardness

    Adds InstanceHardnessCV, which splits folds so hard-to-classify samples are spread evenly rather than clustering in one fold. It is the only new estimator in the window.

    View source ↗
  4. 1y ago

    Metadata routing for samplers and two queued deprecations

    Metadata routing now works for samplers inside a Pipeline, and two deprecations are queued: Pipeline's check_is_fitted warning becomes an error in 0.15, and RUSBoostClassifier's algorithm parameter is on the way out.

    View source ↗
  5. 1y ago

    NumPy 2.0 compatibility

    NumPy 2.0 compatibility in a single line of changelog. Small in content, but it is what let users on the new NumPy keep the package in their stack.

    View source ↗
  6. 2y ago

    scikit-learn 1.5 compatibility release

    A scikit-learn 1.5 compatibility release with no other content. The earliest instance of the pattern that defines this window.

    View source ↗