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 1.1 requires Python 3.8+, numpy 1.17.3+, scipy 1.3.2+ and scikit-learn 1.1-1.3.

Version 1.1.4

wildboar.tree

• Fix An undetected bug which has been around since Version 1.1 has been fixed. The bug affects ShapeletTreeClassifier and ShapeletTreeRegressor and any ensemble using these models. The bug makes it so that the first shapelet is always selected when searching for the best shapelet in a split. In practice this means that any tree fitted with n_shapelet larger than 1 is fitted as if n_shapelet is set to 1 but with the added computational cost. Any results or models constructed using version <=1.1.3 and >= 1.1.0 exhibits the error and the results are most likely worse than expected.

Version 1.1.3

Version 1.1.3 is yet another small maintenance release that adds support for Python 3.12.

Version 1.1.2

Version 1.1.2 is a small maintenance release that fixes interoperability with recent scikit-learn releases and drops the dependency upper bounds. We also include a small number of bug fixes.

• Fix Correctly interoprates with scikit-learn 1.2 and 1.3.

• Fix Drop upper bound on some dependencies and limit the build dependency on cython to <3.0

• Fix Default all wildboar estimators not allow nan.

wildboar.ensemble

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

• Fix Fix a bug where ensemble.ShapeletForestClassifier does not support single dimension 3D input #74.

wildboar.transform

• Fix Correctly set the minimum kernel size to 2 in tree.RocketTransform.

wildboar.tree

• Fix Correctly set min_samples_leaf in tree.RocketTreeClassifier.

• Fix Correctly set the minimum kernel size to 2 in tree.RocketTreeClassifier.

Version 1.1.1

Major changes

Correctly depend on the oldest supported Numpy release.

Changelog

wildboar.annotate

• Fix Correctly return max_motif motifs from annotate.motifs().

wildboar.transform

• Fix Check that transform.SAX and transform.PAA is fitted before transform.

Version 1.1.0

New and changed models

Wildboar 1.1 introduces a large number of new models, explanations and visualizations. Among others we introduce explain.IntervalImportance explain.AmplitudeImportance to highlight important temporal regions; linear_model.RocketClassifier and linear_model.RocketRegressor for state-of-the-art predictive performance using random convoultions. We have also been hard at work adding support for distance.matrix_profile() and implemented motif and regime search in terms of it. We also add support for three new distance metrics, weighted, derivative and weighted derivative dynamic time warping.

Changelog

wildboar.annotate

• Feature annotate.motifs() added to find motifs in time series

• Feature annotate.segment() added to segment time series

wildboar.datasets

• Feature The new module datasets.preprocess for preprocessing datasets.

• Enhancement Setting a global cache directory has been deprecated.

• Enhancement Repository definitions are cached locally to support loading cached datasets while offline.

• Enhancement Support multiple preprocessing directives in datasets.load_dataset().

• API Rename datasets._filter to datasets.filter

wildboar.datasets.outlier

• API Deprecate and remove all labelers.

• Enhancement Add the new function datasets.outlier.minority_outlier().

• Enhancement Add the new function datasets.outlier.majority_outlier().

• Enhancement Add the new function datasets.outlier.kmeans_outlier().

• Enhancement Add the new function datasets.outlier.density_outlier().

• Enhancement Add the new function datasets.outlier.emmott_outlier().

wildboar.distance

• Feature The new distance.matrix_profile() function construct a matrix profile

• Feature Support for Weighted dynamic time warping.

• Feature Support for Derivative (Weighted) dynamic time warping.

• Feature Support for additional Lm distance metrics: minkowski, chebyshev, manhattan.

• Feature Support for cosine an angular distance.

• Feature Support for LCSS, ERP, EDR, MSM, and TWE distance.

• Feature The function distance.dtw.dtw_average() to compute the DTW barycenter average.

• Enhancement The function distance.pairwise_subsequence_distance() generalizes the former distance-function and adds support for n_jobs.

• Enhancement The function distance.paired_subsequence_distance() expands the use-case for the former distance-function.

• Enhancement The function distance.pairwise_distance() adds support for full distance computation.

• Enhancement The function distance.paired_distance() adds support for paired full distance computation.

• Enhancement The function distance.subsequence_match() generalizes the former matches-function and add support for n_jobs.

• Enhancement The function distance.paired_subsequence_match() expands the use-case for the former matches-function.

• Enhancement The function distance.dtw.dtw_alignment() now supports the weight parameter for specifying optional warping weights.

• Feature The new distance.dtw.wdtw_alignment() for weighted DTW alignments.

• API All functions in distance.dtw now require keyword arguments for optional parameters.

• Enhancement All DTW metrics will now default to r=1.0 instead of r=0.0.

wildboar.ensemble

• Feature The new ensemble.RockestRegressor is a classifier that constructs an ensemble of random convolution trees

• Feature The new ensemble.RockestClassifier is a regressor that constructs an ensemble of random convolution trees

• Feature The new ensemble.IntervalForestClassifier is a classifier that constructs an ensemble of interval trees

• Feature The new ensemble.IntervalForestRegressor is a regressor that constructs an ensemble of interval trees

• Feature The new ensemble.PivotForestClassifier is a classifier that constructs an ensemble of pivot trees

• Feature The new ensemble.ProximityForestClassifier is an ensemble of proximity trees.

wildboar.transform

• Feature The new transform.RocketTransform transform time series using random convolutions

• Feature The new transform.RandomShapeletTransform transform time series using the distance to random shapelets

• Feature The new transform.IntervalTransform transform time series using interval features

• Feature The new transform.PivotTransform transform time series using the distance to pivot time series

• Feature The new transform.SAX transforms time series into discretized bins. The implementation is efficient and compress the data to the smallest possible datatype.

• Feature The new transform.PAA transform time series into mean bins.

wildboar.explain

• Feature The new explain.IntervalImportance is an explanation that highlights important regions using permutation importances.

• Feature The new explain.AmplitudeImportance is an explanation that highlights important amplitude regions using SAX and permutation.

• Enhancement explain.counterfactual.ShapeletForestCounterfactual now support multivariate time series.

• Enhancement explain.counterfactual.ShapeletForestCounterfactual now support the verbose argument.

• Enhancement The method argument of explain.counterfactual.counterfactuals() defaults to best. The effect is the same as for infer, which has been deprecated.

• Enhancement explain.counterfactual.counterfactuals() now support a dictionary of method_args.

• API The parameters background_x and background_y of explain.counterfactual.PrototypeCounterfactual has been renamed to train_x and train_y and been deprecated in favor of fit(estimator, x, y).

• API The parameter success to explain.counterfactual.score() is deprecated and has been renamed to valid.

wildboar.linear_model

• Feature The new linear_model.RocketClassifier is a classifier that uses random convolutions and logistic regression

• Feature The new linear_model.RandomShapeletClassifier is a classifier that uses random shapelet distances and logistic regression

• Feature The new linear_model.RocketRegressor is a regressor that uses random convolutions and logistic regression

• Feature The new linear_model.RandomShapeletRegressor is a regressor that uses random shapelet distances and logistic regression

wildboar.tree

• Feature The new tree.IntervalTreeClassifier is a classifier that constructs trees using interval features.

• Feature The new tree.IntervalTreeRegressor is a regressor that constructs trees using interval features.

• Feature The new tree.PivotTreeClassifier is a classifier that uses pivot time series and distance thresholds to construct trees.

• Feature The new tree.ProximityTreeClassifier is a classifier that construct a tree that acts as an ensemble of distance metrics.

• Feature The new tree.RocketTreeClassifier is a classifier that construct random convolution trees

• Feature The new tree.RocketTreeRegressor is a regressor that construct random convolution trees

• Enhancement Add support for dynamically reducing the number of shapelets depending on the currents node depth while constructing tree.ShapeletTreeClassifier and tree.ShapeletTreeRegressor.

• API Rename shapelet of the tree_ attribute to feature.

Other improvements

• We have improved testing, error messages and overall stability.

• Binary builds for Apple ARM processors (M1 and M2)

• Drop support for Python 3.7 and introduce support for Python 3.10