Release history

---

Version 6.1.0

New features

• Support for metadata routing, to pass data such as sample_weight to estimators, scorers, and CV splitters. Read more in the user guide.
• Two new plots: plot_data_splits and plot_cv_splits.
• New set_threshold method to change the threshold of a binary classifier.
• New pos_label attribute to specify the positive label for binary and multilabel classification tasks.

API changes

• The threshold parameter in the evaluate method is deprecated in favour of the set_threshold method.
• Stratification over multiple columns is no longer possible.

Enhancements

• The Imputer class now supports custom strategies for numerical columns by passing a function in place of a strategy name.
• Refactor of the cross-validation splitting strategy.
• Documentation improvements.

Bug fixes

• Fix a bug in conda dependencies for Windows and macOS.

Version 6.0.1

New features

• Support for Python 3.12.

Enhancements

• Packaging license file.

Version 6.0.0

New features

• Completely new module for time series. Read more in the user guide.
• Support for Python 3.11 and drop support for Python 3.8 and Python 3.9.
• New data engines. Read more in the user guide.
• Added the dask parallelization backend.
• Improved memory optimizations. Read more in the user guide.
• Added the iterative strategy for numerical imputation.
• Added the hdbscan strategy to the Pruner class.
• Use the ignore parameter to ignore columns in the dataset.
• New update_traces method to further customize your plots.

API changes

• The plot_results method is divided into plot_results and plot_bootstrap and accepts any metric.
• The FeatureGrouper class no longer accepts a name parameter. Provide the group names directly through the group parameter as dict.
• Rework of the register method.
• The multioutput attribute is deprecated. Multioutput meta-estimators are now assigned automatically.
• Model tags have to be separated from the acronym by an underscore.
• The engine parameter is now a dict.
• The automl method is deprecated.

Enhancements

• Transformations only on y are now accepted, e.g., atom.scale(columns=-1).
• The Imputer class has many more strategies for numerical columns designed for time series.
• The evaluate method highlights the highest score per metric.
• Full support for pandas nullable dtypes.
• The dataset can now be provided as callable.
• The FeatureExtractor class can extract features from the dataset's index.
• Subplots can now share axes on the canvas.
• The save and save_data methods now accept pathlib.Path objects as filename.
• Cleaner representation on hover for the plot_timeline method.
• The cv key in ht_params now accepts a custom cross-validation generator.
• Improved error message for incorrect stratification of multioutput datasets.
• Rework of the shrink method.

Bug fixes

• Fixed a bug where the cross_validate method could fail for pipelines that changed the number of rows.
• Fixed a bug where the Pruner class didn't drop all outlier clusters.
• Fixed a bug where the pipeline could fail for transformers that returned a series.
• Fixed a bug where the pipeline could fail for transformers that reset its internal attributes during fitting.
• Fixed a bug where the register method failed in Databricks.
• Fixed a bug where tuning hyperparameter for a base_estimator inside a custom meta-estimator would fail.
• Fixed a bug where the data properties' @setter could fail for numpy arrays.
• Fixed a bug where reference lines for some plots didn't lie exactly on the unity line.