---
tocdepth: 2
---

(release-notes)=

# Release Notes

## Version 0.15.2 (Feb. 11th, 2026)

A bugfix release that comes with the following changes.

### Fixes

- Fixed an issue that prevented multi-threading from being used by the BOOMER algorithm for updating statistics.
- Fixed an ImportError when using the package "mlrl-testbed" without "mlrl-testbed-slurm" being installed.
- When passing the `--version` or `-v` flag to the command line API, the output is now properly formatted again.

## Version 0.15.1 (Jan. 4th, 2026)

A bugfix release that comes with the following changes.

### Quality-of-Life Improvements

- Pre-built packages are now provided for Windows ARM64 systems.
- In addition to pre-built packages for Linux distributions following the [manylinux](https://packaging.python.org/en/latest/specifications/platform-compatibility-tags/#manylinux) standard, packages are now also provided for [musllinux](https://packaging.python.org/en/latest/specifications/platform-compatibility-tags/#musllinux) systems.

## Version 0.15.0 (Dec. 14th, 2025)

A feature release that comes with the following changes.

```{warning}
This release comes with API changes. For an updated overview of the available parameters and command line arguments, please refer to the [documentation](https://mlrl-boomer.readthedocs.io/en/0.15.0).
```

### Algorithmic Enhancements

- All of our algorithms are now compatible with the scikit-learn API according to the tests provided in the module `sklearn.utils.estimator_checks`. From now on, these tests are applied regularly via continuous integration.

### API Changes

- The Python implementations of algorithms provided by this project do now implement scikit-learn's `decision_function` method, if supported. This method should be used instead of the keyword argument `predict_scores`, which has been removed from the `predict` functions.
- The BOOMER algorithm now uses 32- rather than 64-bit statistics by default. This typically reduces the algorithm's training time and memory footprint without a significant loss of predictive accuracy.
- The package mlrl-testbed-sklearn does now allow to configure and run estimators and meta-estimators provided by the scikit-learn project without the need to write code.
- The package mlrl-testbed-sklearn is now an optional, rather than a mandatory, dependency of the package mlrl-testbed.
- The package mlrl-testbed provides two new modes for re-running experiments and reading their output files, respectively. They can be enabled via the arguments `--mode run` or `--mode read` and are based on reading the metadata.yml files that have been created by previous experiments.
- Experiments are now canceled by default if all of their output files do already exist. This behavior can be adjusted via the newly added argument `--if-outputs-exist`.
- Default arguments can now be defined in the YAML configuration required by mlrl-testbed's batch mode.
- Support for datasets in the LibSVM format has been added to the package mlrl-testbed-sklearn.
- The ranking measures DCG@k and DCG@k with predefined values for k have been added to mlrl-testbed-sklearn.
- The command line argument `--dataset-format` has been added. It allows to explicitly specify the format of a dataset to be loaded.
- The command line argument `--if-input-missing` has been added. It allows to exit the program mlrl-testbed if any error occurs while reading input data. The argument `--exit-on-error` has been replaced with a similar argument `--if-output-error`.
- The command line argument `--print-meta-data` has been added. It allows to print the meta-data of an experiment on the console.
- When using mlrl-testbed's batch mode with the arguments `--separate-folds true` and `--runner slurm`, Slurm jobs that correspond to different folds of a cross validation are now automatically assigned to [job arrays](https://slurm.schedmd.com/job_array.html).
- The command line argument `--wipe-result-dir` has been removed with replacement.
- The option `sparse` has been removed from the command line arguments `--print-label-vectors` and `--save-label-vectors`. Label vectors are now always represented in a sparse format instead of a dense one.
- The option `percentage` has been removed from the command line arguments `--save-evaluation`, `--save-data-characteristics` and `--save-prediction-characteristics`. Values stored via these arguments are now always stored as percentages, if possible.

### Quality-of-Life Improvements

- The implementations of output space statistics used by the BOOMER and SeCo algorithm have been unified and common functionality was moved into the library "libmlrlcommon".

## Version 0.14.2 (Nov. 15th, 2025)

A bugfix release that comes with the following changes.

### Fixes

- Resolved an issue with the YAML schema for validating configuration files required by mlrl-testbed's batch mode.
- If commands in a sbatch script use the bracket notation for specifying options of algorithmic parameters, they are now properly escaped.
- When using mlrl-testbed's read mode with the argument "--separate-folds true" (which is the default) for running experiments for individual cross validation folds, output files are now prevented from being deleted by concurrent experiments.

## Version 0.14.1 (Oct. 14th, 2025)

A bugfix release that comes with the following changes.

### Fixes

- Evaluation results, models and algorithmic parameters are now saved to output files if the argument `--save-all true` is used.
- The number of decimal places used for numerical values written into ARFF files is now chosen more carefully.
- Nominal values are now properly written to ARFF files instead of their indices.
- When writing sparse predictions to ARFF files via the command line argument `--save-predictions true`, their attribute definitions are not malformed anymore.

## Version 0.14.0 (Aug. 22nd, 2025)

A feature release that comes with the following changes.

```{warning}
This release comes with API changes. For an updated overview of the available parameters and command line arguments, please refer to the [documentation](https://mlrl-boomer.readthedocs.io/en/0.14.0).
```

### API Changes

- In batch mode, the argument `--base-dir` is now passed to individual experiments by default.
- The command line argument `--save-meta-data` has been added. It allows to control if meta-data should be written to output files.
- A new package [mlrl-testbed-slurm](https://pypi.org/project/mlrl-testbed-slurm/) has been added to the project. It is an extension that adds support for the Slurm Workload Manager to the package "mlrl-testbed".

### Fixes

- When using the argument `--runner` in batch mode, it is not passed down to the individual experiments anymore.

## Version 0.13.1 (Aug. 4th, 2025)

A bugfix release that comes with the following changes.

### Fixes

- Fixed the release of packages for the Windows platform.

## Version 0.13.0 (Aug. 4th, 2025)

A feature release that comes with the following changes.

```{warning}
This release comes with API changes. For an updated overview of the available parameters and command line arguments, please refer to the [documentation](https://mlrl-boomer.readthedocs.io/en/0.13.0).
```

### API Changes

- The package mlrl-testbed does now provide a batch mode for running multiple experiments at once. It can be enabled via the argument `--mode batch`.
- The command line argument `--sequential-post-optimization true` has been replaced with `--post-optimization sequential`.
- Command line arguments starting with `--store-` have been renamed to use the prefix `--save-`.
- The default value of the command line argument `--save-evaluation` (previously `--store-evaluation`) has been changed from `true` to `false`.
- The command line arguments `--output-dir`, `--create-output-dir` and `--wipe-output-dir` have been renamed to `--result-dir`, `--create-dirs` and `--wipe-result-dir`.
- The command line arguments `--result-dir` (previously `--output-dir`), `--model-load-dir`, `--model-save-dir`, `--parameter-load-dir` and `--parameter-save-dir` do now come with default values.
- The command line arguments `--save-models` and `--save-parameters` have been added for specifying whether models or parameter settings should be written to output files.
- The command line arguments `--load-models` and `--load-parameters` have been added for specifying whether models or parameter settings should be loaded from input files.
- The command line argument `--base-dir` has been added. Relative directories given for the arguments `--result-dir`, `--model-load-dir`, `--model-save-dir`, `--parameter-load-dir` and `--parameter-save-dir` are considered relative to this directory.

### Quality-of-Life Improvements

- Continuous Integration jobs are now used to automatically assign labels and milestones to pull requests.

## Version 0.12.3 (Aug. 2nd, 2025)

A bugfix release that comes with the following changes.

### Fixes

- Fixed another issue that prevented the command line arguments `--print-all` and `--store-all` from working properly.

## Version 0.12.2 (Jul. 31st, 2025)

A bugfix release that comes with the following changes.

### Quality-of-Life Improvements

- The static type checker [mypy](https://www.mypy-lang.org/) is now applied to the project's Python code via Continuous Integration.

### Fixes

- Fixed an issue that prevented the command line arguments `--print-all` and `--store-all` from working properly.

## Version 0.12.1 (Jul. 6th, 2025)

A bugfix release that comes with the following changes.

### Quality-of-Life Improvements

- For testing Python packages under more realistic conditions, they are now installed into isolated CI jobs, separated from the build jobs.
- [pytest](https://pytest.org/) is now used for testing the Python code. This enables, among other things, to use parameterized tests.
- Python tests are now divided into blocks that are run in parallel by CI jobs

### Fixes

- Fixed an issue that caused the options `first_fold` and `last_fold` to be swapped when using the command line argument `--data-split cross-validation`.

## Version 0.12.0 (Jun. 29th, 2025)

A feature release that comes with the following changes.

```{warning}
This release comes with API changes. For an updated overview of the available parameters and command line arguments, please refer to the [documentation](https://mlrl-boomer.readthedocs.io/en/0.12.0).
```

### API Changes

- The executable `testbed` has been renamed to `mlrl-testbed`.
- The command line argument `--model-dir` has been replaced with the arguments `--model-load-dir` and `--model-save-dir`, which enables to save models to a different directory than the one they are loaded from.
- The command line argument `--parameter-dir` has been replaced with the arguments `--parameter-load-dir` and `--parameter-save-dir`. The former specifies the directory, parameter settings should be loaded from, whereas the latter replaces the argument `--store-parameters`.
- The command line argument `--store-predictions` does now write ARFF files where the ground truth is replaced with the predictions of a model. Accordingly, the ARFF files written via the new argument `--store-ground-truth` contains the original ground truth.
- The command line argument `--evaluate-training-data` has been renamed to `--predict-for-training-data`. Analogously, a new argument `--predict-for-test-data` has been added.
- By default, the command line API is not terminated anymore when an error occurs while writing output data. This behavior can be changed via the new argument `--exit-on-error`.
- When passing the value `cross-validation` to the command line argument `--data-split`, the options `first_fold` and `last_fold` can now be used to specify a range of folds to be run. The option `current_fold` has been removed.
- Output directories are now automatically created by the command line API. This behavior can be disabled via the newly added argument `--create-output-dir`.
- The command line argument `--wipe-output-dir` has been added. It allows to prevent existing output files from being deleted before an experiment starts.
- The new command line arguments `--store-all` and `--print-all` allow to print all output data on the console or to write it to files.
- The options `min_samples` and `max_samples` have been added to the values of the command line arguments `--feature-sampling` and `--instance-sampling`.
- The indices of nominal and ordinal features are now provided to a learner's `fit`-method via the keyword arguments `nominal_feature_indices` and `ordinal_feature_indices`.
- The Python API does now allow to provide custom weights for training examples to a learner's `fit`-method via the keyword argument `sample_weights`.
- The Python API of the class `RuleModel` does now provide an easy way to access the rules in the model.

### Algorithmic Enhancements

- The BOOMER algorithm can now be configured to use either 32- or 64-bit floating point values for gradients and Hessians via the command line argument `--statistic-type`. Using lower-precision values might speed up training at the risk of losing training accuracy.
- Efficient data types and data structures are now used for storing binary scores calculated by the SeCo algorithm.
- Unnecessary conversions from integer weights to floating point values are now avoided.

### Fixes

- Correct data types are now used for ground truth matrices when using the command line argument `--problem-type regression`. Previously, the values in these matrices were completely off, rendering any experiments using them invalid.
- An issue that caused thresholds and probabilities of isotonic regression models being swapped in output files when using the command line arguments `--store-marginal-probability-calibration-model` and `--store-joint-probability-calibration-model` has been fixed.

### Quality-of-Life Improvements

- The Python package "mlrl-util" has been added. It provides common functionalities of the packages "mlrl-common" and "mlrl-testbed".
- The Python package "mlrl-testbed-arff" and "mlrl-testbed-sklearn" have been added as extensions to the package "mlrl-testbed".
- The Python package "mlrl-testbed" has completely been refactored, establishing it as a standalone package, independent of the package "mlrl-common".
- The Python packages "mlrl-common", "mlrl-seco" and "mlrl-boosting" have been restructured by introducing submodules.
- C++ 20 is now required for compiling the project.
- The integration tests do now check the contents of output files.
- The build targets `format_python` and `test_format_python` now employ [autoflake](https://github.com/PyCQA/autoflake) to detect and remove unused variables and imports, as well as unnecessary `pass` statements from Python and Cython source files.
- The build targets `format_cfg` and `test_format_cfg` have been added. They enforce a consistent style for .cfg files by employing the package [config-formatter](https://github.com/Delgan/config-formatter).
- The tool [cython-lint](https://github.com/MarcoGorelli/cython-lint) is now applied to Cython source files via Continuous Integration.

## Version 0.11.4 (Feb. 27th, 2025)

A bugfix release that comes with the following changes.

### Fixes

- The parameter `--head-type auto` can now be set via the command line API without an error being raised.

## Version 0.11.3 (Jan. 30th, 2025)

A bugfix release that comes with the following changes.

### Fixes

- Fixed faulty calculation of gradients and Hessians when using the BOOMER regression algorithm with the command line arguments `--loss squared-error-non-decomposable` and `--head-type single`.

### Quality-of-Life Improvements

- Linux ARM64 packages are now build natively instead of using virtualization, which speeds up the build process tremendously.
- Continuous Integration is now used to automatically update the versions of GitHub runners on a regular basis.

## Version 0.11.2 (Jan. 23rd, 2025)

A bugfix release that comes with the following changes.

### Fixes

- When building macOS packages, we do now manually compile OpenMP to ensure that the bundled library matches the target platform.

### Quality-of-Life Improvements

- The build system now uses a lightweight custom implementation instead of [SCons](https://scons.org/) and is better modularized to avoid unnecessary runs of Continuous Integration jobs when only certain parts of it are modified.
- Releases are now automated via Continuous Integration, including the update of the project's changelog.
- The presentation of algorithmic parameters in the documentation has been improved.
- Continuous Integration is now used to automatically update outdated GitHub Actions, outdated Python dependencies, and the Doxygen configuration on a regular basis.
- Continuous Integration is now used to check the documentation for broken links and spelling mistakes.
- The syntax of TOML files is now validated by applying [taplo](https://github.com/tamasfe/taplo) via Continuous Integration.

## Version 0.11.1 (Sep. 24th, 2024)

A bugfix release that fixes the following issues:

- A crash has been fixed that could occur when using the command line argument `--print-rules` or `--store-rules` with a dataset that contains nominal features.

## Version 0.11.0 (Aug. 10th, 2024)

A major update to the BOOMER algorithm that introduces the following changes.

```{warning}
This release comes with several API changes. For an updated overview of the available parameters and command line arguments, please refer to the [documentation](https://mlrl-boomer.readthedocs.io/en/0.11.0/).
```

### Algorithmic Enhancements

- **The BOOMER algorithm can be used for solving regression problems**, including single- and multi-output regression problems.

### Additions to the Command Line API

- **Custom algorithms can now be easily integrated** with the command line API due to the ability to dynamically load code from your own Python modules or source files, as illustrated [here](https://mlrl-boomer.readthedocs.io/en/0.11.0/user_guide/testbed/runnables.html)
- **The value to be used for sparse elements in the feature matrix can be specified** via the argument `--sparse-feature-value`.

### API Changes

- The Python module or source file providing an integration with the machine learning algorithm to be used by the command line API must now be specified as described [here](https://mlrl-boomer.readthedocs.io/en/0.11.0/user_guide/testbed/arguments.html#basic-usage).
- Several parameters and their values have been renamed to better reflect the scope of the project, which now includes multi-output regression problems. For an up-to-date list of parameters, please refer to the [documentation](https://mlrl-boomer.readthedocs.io/en/0.11.0/).
- Rules with complete heads are now learned by default when using a decomposable loss function and a dense format for storing statistics.

## Version 0.10.2 (Aug. 9th, 2024)

A bugfix release that fixes the following issues.

- A rare issue that may result in rules with empty heads being learned when using the argument `--head-type partial-dynamic` has been fixed.

## Version 0.10.1 (Aug. 1st, 2024)

A bugfix release that comes with the following changes.

- If the sparse value of a feature matrix is provided to the Python API, it is now properly taken into account when converting into a dense matrix.
- The C++ code is now checked for common issues by applying `cpplint` via Continuous Integration.
- The styling of YAML files is now verified by applying `yamlfix` via Continuous Integration.

## Version 0.10.0 (May 5th, 2024)

A major update to the BOOMER algorithm that introduces the following changes.

```{warning}
This release comes with several API changes. For an updated overview of the available parameters and command line arguments, please refer to the [documentation](https://mlrl-boomer.readthedocs.io/en/0.10.0/).
```

### Algorithmic Enhancements

- **The project does now provide a Separate-and-Conquer (SeCo) algorithm** based on traditional rule learning techniques that are particularly well-suited for learning interpretable models.
- **Space-efficient data structures are now used for storing feature values**, depending on whether the feature is numerical, ordinal, nominal, or binary. This also enables to use optimized code paths for dealing with these different types of features.
- **The implementation of feature binning has been reworked** in a way that avoids redundant code and results in a reduction of training times due to the use of the data structures mentioned above.
- **The value to be used for sparse elements of a feature matrix** can now be specified via the C++ or Python API.
- **Nominal and ordinal feature values are now represented as integers** to avoid issues due to limited floating point precision.
- **Safe comparisons of floating point values** are now used to avoid issues due to limited floating point precision.
- **Fundamental data structures for vectors and matrices have been reworked** to ease reusing existing functionality and avoiding redundant code.

### Additions to the Command Line API

- **Information about the program can now be printed** via the argument `-v` or `--version`.
- **Data characteristics do now include the number of ordinal features** when printed on the console or written to a file via the command line argument `--print-data-characteristics` or `--store-data-characteristics`.

### Bugfixes

- An issue has been fixed that caused the number of numerical and nominal features to be swapped when using the command line arguments `--print-data-characteristics` or `--store-data-characteristics`.
- The correct directory is now used for loading and saving parameter settings when using the command line arguments `--parameter-dir` and `--store-parameters`.

### API Changes

- The option `num_threads` of the parameters `--parallel-rule-refinement`, `--parallel-statistic-update` and `--parallel-prediction` has been renamed to `num_preferred_threads`.

### Quality-of-Life Improvements

- The documentation has been updated to a more modern theme supporting light and dark theme variants.
- A build option that allows disabling multi-threading support via OpenMP at compile-time has been added.
- The groundwork for GPU support was laid. It can be disabled at compile-time via a build option.
- Added support for unit testing the project's C++ code. Compilation of the tests can be disabled via a build option.
- The Python code is now checked for common issues by applying `pylint` via Continuous Integration.
- The Makefile has been replaced with wrapper scripts triggering a [SCons](https://scons.org/) build.
- Development versions of wheel packages are now regularly built via Continuous Integration, uploaded as artifacts, and published on [Test-PyPI](https://test.pypi.org/).
- Continuous integration is now used to maintain separate branches for major, feature, and bugfix releases and keep them up-to-date.
- The runtime of Continuous Integration jobs has been optimized by running individual steps only if necessary, caching files across subsequent runs, and making use of parallelization.
- When tests are run via Continuous Integration, a summary of the test results is now added to pull requests and GitHub workflows.
- Markdown files are now used for writing the documentation.
- A consistent style is now enforced for Markdown files by applying the tool `mdformat` via Continuous Integration.
- C++ 17 or newer is now required for compiling the project.

## Version 0.9.0 (Jul. 2nd, 2023)

A major update to the BOOMER algorithm that introduces the following changes.

```{warning}
This release comes with several API changes. For an updated overview of the available parameters and command line arguments, please refer to the [documentation](https://mlrl-boomer.readthedocs.io/en/0.9.0/).
```

### Algorithmic Enhancements

- **Sparse matrices can now be used to store gradients and Hessians** if supported by the loss function. The desired behavior can be specified via a new parameter `--statistic-format`.
- **Rules with partial heads can now be learned** by setting the parameter `--head-type` to the value `partial-fixed`, if the number of predicted labels should be predefined, or `partial-dynamic`, if the subset of predicted labels should be determined dynamically.
- **A beam search can now be used** for the induction of individual rules by setting the parameter `--rule-induction` to the value `top-down-beam-search`.
- **Variants of the squared error loss and squared hinge loss**, which take all labels of an example into account at the same time, can now be used by setting the parameter `--loss` to the value `squared-error-example-wise` or `squared-hinge-example-wise`.
- **Probability estimates can be obtained for each label independently or via marginalization** over the label vectors encountered in the training data by setting the new parameter `--probability-predictor` to the value `label-wise` or `marginalized`.
- **Predictions that maximize the example-wise F1-measure can now be obtained** by setting the parameter `--classification-predictor` to the value `gfm`.
- **Binary predictions can now be derived from probability estimates** by specifying the new option `based_on_probabilities`.
- **Isotonic regression models can now be used** to calibrate marginal and joint probabilities predicted by a model via the new parameters `--marginal-probability-calibration` and `--joint-probability-calibration`.
- **The rules in a previously learned model can now be post-optimized** by reconstructing each one of them in the context of the other rules via the new parameter `--sequential-post-optimization`.
- **Early stopping or post-pruning can now be used** by setting the new parameter `--global-pruning` to the value `pre-pruning` or `post-pruning`.
- **Single labels can now be sampled in a round-robin fashion** by setting the parameter `--feature-sampling` to the new value `round-robin`.
- **A fixed number of trailing features can now be retained** when the parameter `--feature-sampling` is set to the value `without-replacement` by specifying the option `num_retained`.

### Additions to the Command Line API

- **Datasets in the MEKA format are now supported.**
- **Certain characteristics of binary predictions can be printed or written to output files** via the new arguments `--print-prediction-characteristics` and `--store-prediction-characteristics`.
- **Unique label vectors contained in the training data can be printed or written to output files** via the new arguments `--print-label-vectors` and `--store-label-vectors`.
- **Models for the calibration of marginal or joint probabilities can be printed or written to output files** via the new arguments `--print-marginal-probability-calibration-model`, `--store-marginal-probability-calibration-model`, `--print-joint-probability-calibration-model` and `--store-joint-probability-calibration-model`.
- **Models can now be evaluated repeatedly, using a subset of their rules with increasing size,** by specifying the argument `--incremental-prediction`.
- **More control of how data is split into training and test sets** is now provided by the argument `--data-split` that replaces the arguments `--folds` and `--current-fold`.
- **Binary labels, scores, or probabilities can now be predicted,** depending on the value of the new argument `--prediction-type`, which can be set to the values `binary`, `scores`, or `probabilities`.
- **Individual evaluation measures can now be enabled or disabled** via additional options that have been added to the arguments `--print-evaluation` and `--store-evaluation`.
- **The presentation of values printed on the console has vastly been improved.** In addition, options for controlling the presentation of values to be printed or written to output files have been added to various command line arguments.

### Bugfixes

- The behavior of the parameter `--label-format` has been fixed when set to the value `auto`.
- The behavior of the parameters `--holdout` and `--instance-sampling` has been fixed when set to the value `stratified-label-wise`.
- The behavior of the parameter `--binary-predictor` has been fixed when set to the value `example-wise` and using a model that has been loaded from disk.
- Rules are now guaranteed to not cover more examples than specified via the option `min_coverage`. The option is now also taken into account when using feature binning. Alternatively, the minimum coverage of rules can now also be specified as a fraction via the option `min_support`.

### API Changes

- The parameter `--early-stopping` has been replaced with a new parameter `--global-pruning`.
- The parameter `--pruning` has been renamed to `--rule-pruning`.
- The parameter `--classification-predictor` has been renamed to `--binary-predictor`.
- The command line argument `--predict-probabilities` has been replaced with a new argument `--prediction-type`.
- The command line argument `--predicted-label-format` has been renamed to `--prediction-format`.

### Quality-of-Life Improvements

- Continuous integration is now used to test the most common functionalities of the BOOMER algorithm and the corresponding command line API.
- Successful generation of the documentation is now tested via Continuous Integration.
- Style definitions for Python and C++ code are now enforced by applying the tools `clang-format`, `yapf`, and `isort` via Continuous Integration.

## Version 0.8.2 (Apr. 11th, 2022)

A bugfix release that solves the following issues:

- Fixed prebuilt packages available at [PyPI](https://pypi.org/project/mlrl-boomer/).
- Fixed output of nominal values when using the option `--print-rules true`.

## Version 0.8.1 (Mar. 4th, 2022)

A bugfix release that solves the following issues:

- Missing feature values are now dealt with correctly when using feature binning.
- A rare issue that may cause segmentation faults when using instance sampling has been fixed.

## Version 0.8.0 (Jan. 31, 2022)

A major update to the BOOMER algorithm that introduces the following changes.

```{warning}
This release comes with changes to the command line API. For an updated overview of the available parameters, please refer to the [documentation](https://mlrl-boomer.readthedocs.io/en/0.8.0/).
```

- The programmatic C++ API was redesigned for a more convenient configuration of algorithms. This does also drastically reduce the amount of wrapper code that is necessary to access the API from other programming languages and therefore facilitates the support of additional languages in the future.
- An issue that may cause segmentation faults when using stratified sampling methods for the creation of holdout sets has been fixed.
- Pre-built packages for Windows systems are now available at [PyPI](https://pypi.org/project/mlrl-boomer/).
- Pre-built packages for Linux ARM64 systems are now provided.

## Version 0.7.1 (Dec. 15, 2021)

A bugfix release that solves the following issues:

- Fixes an issue preventing the use of dense representations of ground truth label matrices that was introduced in version 0.7.0.
- Pre-built packages for macOS systems are now available at [PyPI](https://pypi.org/project/mlrl-boomer/).
- Linux and macOS packages for Python 3.10 are now provided.

## Version 0.7.0 (Dec. 5, 2021)

A major update to the BOOMER algorithm that introduces the following changes:

- L1 regularization can now be used.
- A more space-efficient data structure is now used for the sparse representation of binary predictions.
- The Python API does now allow to access the rules in a model in a programmatic way.
- It is now possible to output certain characteristics of training datasets and rule models.
- Pre-built packages for the Linux platform are now available at [PyPI](https://pypi.org/project/mlrl-boomer/).
- The [documentation](https://mlrl-boomer.readthedocs.io) has vastly been improved.

## Version 0.6.2 (Oct. 4, 2021)

A bugfix release that solves the following issues:

- Fixes a segmentation fault when a sparse feature matrix should be used for prediction that was introduced in version 0.6.0.

## Version 0.6.1 (Sep. 30, 2021)

A bugfix release that solves the following issues:

- Fixes a mathematical problem when calculating the quality of potential single-label rules that was introduced in version 0.6.0.

## Version 0.6.0 (Sep. 6, 2021)

A major update to the BOOMER algorithm that introduces the following changes.

```{warning}
This release comes with changes to the command line API. For brevity and consistency, some parameters and/or their values have been renamed. Moreover, some parameters have been updated to use more reasonable default values. For an updated overview of the available parameters, please refer to the [documentation](https://mlrl-boomer.readthedocs.io/en/0.6.0/).
```

- The parameter `--instance-sampling` does now allow to use stratified sampling (`stratified-label-wise` and `stratified-example-wise`).
- The parameter `--holdout` does now allow to use stratified sampling (`stratified-label-wise` and `stratified-example-wise`).
- The parameter `--recalculate-predictions` does now allow to specify whether the predictions of rules should be recalculated on the entire training data, if instance sampling is used.
- An additional parameter (`--prediction-format`) that allows to specify whether predictions should be stored using dense or sparse matrices has been added.
- The code for the construction of rule heads has been reworked, resulting in minor performance improvements.
- The unnecessary calculation of Hessians is now avoided when used single-label rules for the minimization of a non-decomposable loss function, resulting in a significant performance improvement.
- A programmatic C++ API for configuring algorithms, including the validation of parameters, is now provided.
- A documentation is now available [online](https://mlrl-boomer.readthedocs.io).

## Version 0.5.0 (Jun. 27, 2021)

A major update to the BOOMER algorithm that introduces the following changes:

- Gradient-based label binning (GBLB) can be used to assign labels to a predefined number of bins.

## Version 0.4.0 (Mar. 31, 2021)

A major update to the BOOMER algorithm that introduces the following changes:

- Large parts of the code have been refactored, and the core algorithm has been migrated to C++ entirely. It is now built and compiled using Meson and Ninja, which results in drastically reduced compile times.
- The (label- and example-wise) logistic loss functions have been rewritten to better prevent numerical problems.
- Approximate methods for evaluating potential conditions of rules, based on unsupervised binning methods (currently equal-width- and equal-frequency-binning), have been added.
- The parameter `--predictor` does now allow using different algorithms for prediction (`label-wise` or `example-wise`).
- An early stopping mechanism has been added, which allows to stop the induction of rules as soon as the quality of the model does not improve on a holdout set.
- Multi-threading can be used to parallelize the prediction for different examples across multiple CPU cores.
- Multi-threading can be used to parallelize the calculation of gradients and Hessians for different examples across multiple CPU cores.
- Probability estimates can be predicted when using the loss function `label-wise-logistic-loss`.
- The algorithm does now support datasets with missing feature values.
- The loss function `label-wise-squared-hinge-loss` has been added.
- Experiments using single-label datasets are now supported out of the box.

## Version 0.3.0 (Sep. 14, 2020)

A major update to the BOOMER algorithm that features the following changes:

- Large parts of the code (loss functions, calculation of gradients/Hessians, calculation of predictions/quality scores) have been refactored and rewritten in C++. This comes with a constant speed-up of training times.
- Multi-threading can be used to parallelize the evaluation of a rule's potential refinements across multiple CPU cores.
- Sparse ground truth label matrices can now be used for training, which may reduce the memory footprint in case of large datasets.
- Additional parameters (`feature-format` and `label-format`) that allow to specify the preferred format of the feature and label matrix have been added.

## Version 0.2.0 (Jun. 28, 2020)

A major update to the BOOMER algorithm that features the following changes:

- Includes many refactorings and quality of live improvements. Code that is not directly related with the algorithm, such as the implementation of baselines, has been removed.
- The algorithm is now able to natively handle nominal features without the need for pre-processing techniques such as one-hot encoding.
- Sparse feature matrices can now be used for training and prediction, which reduces the memory footprint and results in a significant speed-up of training times on some datasets.
- Additional hyperparameters (`min_coverage`, `max_conditions` and `max_head_refinements`) that provide fine-grained control over the specificity/generality of rules have been added.

## Version 0.1.0 (Jun. 22, 2020)

The first version of the BOOMER algorithm used in the following publication:

*Michael Rapp, Eneldo Loza Mencía, Johannes Fürnkranz and Eyke Hüllermeier. Gradient-based Label Binning in Multi-label Classification. In: Proceedings of the European Conference on Machine Learning and Knowledge Discovery in Databases (ECML-PKDD), 2021, Springer.*

This version supports the following features to learn an ensemble of boosted classification rules:

- Different label-wise or example-wise loss functions can be minimized during training (optionally using L2 regularization).
- The rules may predict for a single label, or for all labels (which enables to model local label dependencies).
- When learning a new rule, random samples of the training examples, features or labels may be used, including different techniques such as sampling with or without replacement.
- The impact of individual rules on the ensemble can be controlled using shrinkage.
- The conditions of a recently induced rule can be pruned based on a hold-out set.
- The algorithm currently only supports numerical or ordinal features. Nominal features can be handled by using one-hot encoding.