wildboar

A new release with multiple new estimators and improvements!

See full change log at the website.

This release fixes a serious bug in ShapeletForestClassifier and ShapeletTreeClassifier where trees would only be constructed using a single random shapelet. If the model is used the performance is likely worse than expected.

Read full changelog at the website.

Small maintenance release with support for Python 3.12.

This release fixes several (minor) bugs and enables support for scikit-learn 1.2 and 1.3.

Read full changelog at the website.

This release fixes the numpy install dependency of v1.1 and sets it to the oldest supported dumpy version (1.17.3).

See full change log at the website.

This release contains multiple breaking changes.

See full change log at the website.

Fixes

• Fix #25 for multivariate time series

Fixed

• Fix bug #25 which is present if sklearn >= 1.0 is installed

Fixes

• Fix a serious issue in tree._tree_builder.Tree where platform dependent
  integer sizes were used (see #22).

Fixes

• Binary builds

Fixes

• Binary builds for GNU/Linux

Fixes

• Fix default repository endpoint. Previous releases are now deprecated.

Added

• Add model_selection.outlier.RepeatedOutlierSplit to cross-validate
  outlier detection algorithms

Fixed

• Fix bug in filter of datasets.load_datasets
• Fix the number of outliers when setting n_outliers to None for
  KMeansLabeler
• Fix the number of outliers when setting n_outliers to float for
  MinorityLabeler

Added

• Add filter keyword-argument to datasets.load_datasets

Fixed

• Fix bug in source distribution

Added

• Add MajorityLabeler to construct synthetic outlier datasets

Fixed

• Fix bug in IsolationShapeletForest where max_depth was incorrectly computed

Added

• Model agnostic counterfactual explanations has been added.
• Shapelet forest counterfactual explanations has been refined.
• KNearestNeighbors counterfactual explanations has been refined.
• Synthetic generation of outlier detection datasets.
• IsolationShapeletForest has been added. A novel semi-supervised method for detecting
  time series outliers.
• Fast computation of scaled and unscaled dynamic time warping (using the UCRSuite algorithm).
• LB_Keogh lower bound and envelope.
• Add new class Repository which represents a collection of bundles
• datasets.set_cache_dir to globally change the default cache directory
• datasets.clear_cache to clear the cache

Deprecated

• datasets.load_all_datasets has been replaced by load_datasets

Changed

• wildboar.datasets.install_repository now installs a repository instead of a bundle
• Rename Repository to Bundle

Added

• Added a counterfactual explainability module

Changed

• Complete rewrite of the shapelet tree representation to allow releasing GIL.
  The prediction of trees should be backwards compatible, i.e., trees fitted using
  the new versions are functionally equivalent to the old but with another internal
  representation.

Fixed

• Improved caching of lower-bound for DTW
  The DTW subsequence search implementation has been improved by caching
  DTW lower-bound information for repeated calls with the same
  subsequece. This slightly increases the memory requirement, but can
  give significantly improved performance under certain circumstances.

• Allow shapelet information to be extracted
  A new attribute ts_info is added to Shapelet (which is accessible
  from tree.root_node_.shapelet). ts_info returns a tuple
  (ts_index, ts_start, length) with information about the index (in
  the x used to fit, fit(x, y), the model) and the start position of
  the shapelet. For a shapelet tree/forest fit on x the shapelet in a
  particular node is given by x[ts_index, ts_start:(ts_start + length)].