What’s new

• Feature: something that you couldn’t do before.

• Efficiency: an existing feature now may not require as much computation or memory.

• Enhancement: a miscellaneous minor improvement.

• Fix: something that previously didn’t work as documentated

• API: you will need to change your code to have the same effect in the future; or a feature will be removed in the future.

Dependencies

Wildboar 1.2 requires Python 3.8+, numpy 1.19.5+, scipy 1.6.0+ and scikit-learn 1.3+.

Version 1.2.1

Limit scikit-learn to a version before 1.6 and build for Python 3.13.

Version 1.2.0

New and changed models

Wildboar 1.2 introduces several new models.

• transform.HydraTransform: a new convolution based dictionary transformation method as described by Dempster et al., (2023).

• linear_model.HydraClassifier: a new convolution based dictionary classifier as described by Dempster et al., (2023).

• distance.KNeighborsClassifier: the traditional k-neighbors classifier using wildboar native distance metrics, including the full suite of optimized elastic metrics.

• distance.KMeans: the traditional k-means clustering algorithm. Compared to scikit-learn, this implementation supports dtw and wdtw.

• distance.KMedoids: the traditional k-medoids clustering algorithm with support for all elastic metrics.

• ensemble.ElasticEnsembleClassifier: the elastic ensemble classifier as described by Lines and Bagnall (2015).

• transform.DilatedShapeletTransform: a new shapelet based transform as described by Guillaume et al., (2022).

• linear_model.DilatedShapeletClassifier: a new shapelet based classifier as described by Guillaume et al., (2022).

• transform.CastorTransform: a new shapelet based transform using competing shapelets introduced in Samsten and Lee (2024).

• linear_model.CastorClassifier: a new shapelet based classifier using competing shapelets introduced in Samsten and Lee (2024).

Changelog

wildboar.datasets

• API Drop support for specifying a dataset version in load_dataset. Support was dropped to ensure consistency between the repository declaration of arrays and what is available in the downloaded bundle.

• Fix Correctly detect duplicate repositories.

• Fix Defer repository refresh to first use.

wildboar.distance

• Enhancement Improve support for 3darrays in distance.pairwise_distance and distance.paired_distance. By setting dim='mean', the mean distance over all dimensions are computed and by setting dim='full' the distance over all dimensions are returned. The default value for dim will change to “mean” in 1.3. For 3darrays, we issue a deprecation warning for the current default value.

• Enhancement Add support for standardizing all subsequence metrics. Prefix the name of the metric with "scaled_" to use the standardized metric, e.g., “scaled_euclidean” or “scaled_msm”.

• Enhancement Add support for callable metrics. To support standardizing, we introduce new keyword parameter that pairs with the “metric” parameter called scale that, if set to True scale all subsequences before applying the metric. The effect of setting scale=True is the same as passing a scaled metric, e.g., “scaled_euclidean”.

• Feature A new function distance.argmin_distance which takes as input two arrays X and Y and finds, for each sample in X, the indices of the k samples in Y with the smallest distance to the i:th sample in X.

• Feature A new function distance.distance_profile which takes a subsequence Y and a (collection) of time series X and returns the distance from Y to all subsequences of the same length of the i:th sample in X, with support for dilation and padding.

• Feature A new function distance.argmin_subsequence_distance which takes two paired arrays of subsequences and samples and finds the k smallest matching positions for each sample/subsequence pair.

• Feature Enables support for n_jobs in distance.pairwise_distance, distance.paired_distance.

• Feature Add support for Amercing Dynamic Time Warping (subsequence) distance.

• Feature Add support for LCSS subsequence distance.

• Feature Add support for EDR subsequence distance.

• Feature Add support for TWE subsequence distance.

• Feature Add support for MSM subsequence distance.

• Feature Add support for ERP subsequence distance.

• Fix Fix a bug in angular distance leading to NaN values.

• Fix Fix a bug in angular distance subsequence matching where an incorrect threshold was set.

• Fix Fix the return value of distance.paired_distance to (n_dims, n_samples) when dim=”full”.

• API Rename LCSS threshold parameter to epsilon. We will remove threshold in 1.4.

• API Rename EDR threshold parameter to epsilon. We will remove threshold in 1.4.

• API Rename _distance.DistanceMeasure to Metric and _distance.SubsequenceDistanceMeasure to SubsequenceMetric. The change only affect code that cimport modules.

• API The default value of threshold in distance.subsequence_match has changed to “auto”. The old value “best” has been deprecated and will be removed in 1.3.

wildboar.ensemble

• Feature Add support for multiple metrics in ensemble.ShapeletForestClassifier, ensemble.ShapeletForestRegressor. All estimators with a metric parameter and which implements the ShapeletMixin are affected by this change.

• API Rename the constructor parameter base_estimator to estimator in ensemble.BaggingClassifier and ensemble.BaggingRegressor. base_estimator is deprecated in 1.2 and will be removed in 1.4.

• API Change the tuple argument for kernel_size to two new parameters min_size and max_size. This change affect tree.RocketForestClassifier and tree.RocketForestRegressor.

• Fix Fix a bug where sampling was incorrectly set for ensemble.RocketForestClassifier and ensemble.RocketForestRegressor.

• API Change the default value of n_shapelets to “log2” for ensemble.ShapeletForestClassifier and ensemble.ShapeletForestRegressor.

• API Drop support for criterion="mse" in ensemble.ShapeletForestRegressor and ensemble.ExtraShapeletTreesRegressor.

wildboar.explain.counterfactual

• Feature Add support for KNeighborsClassifiers fitted with any metric in explain.counterfactual.KNeighborsCounterfactual. We allow for using different methods for finding the counterfactuals for n_neighbors > 1 by setting method=’mean’ or method=’medoid’. We have also improved the way in which cluster centroids are selected, resulting in a more robust counterfactuals.

wildboar.linear_model

• API Undeprecate the normalize parameter from linear_model.RocketClassifier and linear_model.RocketRegressor.

wildboar.transform

• Feature Add support for multiple metrics in transform.RandomShapeletTransform by passing a list of metric specifications. See the documentation for details.

• Enhancement Rename the parameter value log for the parameter n_intervals in transform.IntervalTransform to log2. The old value is deprecated and will be removed in 1.4.

• Feature Improve the metric specification for transform.PivotTransform.

• API Change the tuple argument for kernel_size to two new parameters min_size and max_size. This change affect transform.RocketTransform.

wildboar.tree

• Feature Add support for multiple metrics in tree.ShapeletTreeClassifier, tree.ShapeletTreeRegressor. All estimators with a metric parameter and which implements the ShapeletMixin is affected by this change.

• Fix Correctly use MSM distance measure in tree.ProximityTreeClassifier.

• Fix Correctly set min_samples_leaf in tree.RocketTreeClassifier and RocketTreeRegressor.

• API Change the tuple argument for kernel_size to two new parameters min_size and max_size. This change affect tree.RocketTreeClassifier and tree.RocketTreeRegressor.

• API The metric_factories parameter of tree.ProximityTreeClassifier has been renamed to metric. We have deprecated metric_factories and it will be removed in 1.4. We also introduce the metric_params argument for single metric uses.

• API Change the default value of n_shapelets to “log2” for tree.ShapeletTreeClassifier and tree.ShapeletTreeRegressor.

• API Drop support for criterion="mse" in tree.ShapeletTreeRegressor and tree.ExtraShapeletTreeRegressor.

Other improvements

• Remove all dependencies on deprecated Numpy APIs.

• Migrate to the new scikit-learn parameter validation framework.