mlrl-boomer 0.15.0

A scikit-learn implementation of BOOMER - an algorithm for learning gradient boosted multi-label output rules

🔗 Important links: Documentation | Issue Tracker | Changelog | License

This software package provides the official implementation of BOOMER - an algorithm for learning gradient boosted multi-output rules that uses gradient boosting for learning an ensemble of rules that is built with respect to a specific multivariate loss function. It integrates with the popular scikit-learn machine learning framework.

The problem domains addressed by this algorithm include the following:

• Multi-label classification: The goal of multi-label classification is the automatic assignment of sets of labels to individual data points, for example, the annotation of text documents with topics.
• Multi-output regression: Multivariate regression problems require to predict for more than a single numerical output variable.

The BOOMER Algorithm

To provide a versatile tool for different use cases, great emphasis is put on the efficiency of the implementation. Moreover, to ensure its flexibility, it is designed in a modular fashion and can therefore easily be adjusted to different requirements. This modular approach enables implementing different kind of rule learning algorithms (see packages mlrl-common and mlrl-seco).

📖 References

The algorithm was first published in the following paper. A preprint version is publicly available here.

Michael Rapp, Eneldo Loza Mencía, Johannes Fürnkranz Vu-Linh Nguyen and Eyke Hüllermeier. Learning Gradient Boosted Multi-label Classification Rules. In: Proceedings of the European Conference on Machine Learning and Knowledge Discovery in Databases (ECML-PKDD), 2020, Springer.

If you use the algorithm in a scientific publication, we would appreciate citations to the mentioned paper.

🔧 Functionalities

The algorithm that is provided by this project currently supports the following core functionalities for learning ensembles of boosted classification or regression rules.

Deliberate Loss Optimization

• Decomposable or non-decomposable loss functions can be optimized in expectation.
• L1 and L2 regularization can be used.
• Shrinkage (a.k.a. the learning rate) can be adjusted for controlling the impact of individual rules on the overall ensemble.

Different Prediction Strategies

• Various strategies for predicting scores, binary labels or probabilities are available, depending on whether a classification or regression model is used.
• Isotonic regression models can be used to calibrate marginal and joint probabilities predicted by a classification model.

Flexible Handling of Input Data

• Native support for numerical, ordinal, and nominal features eliminates the need for pre-processing techniques such as one-hot encoding.
• Handling of missing feature values, i.e., occurrences of NaN in the feature matrix, is implemented by the algorithm.

Fine-grained Control over Model Characteristics

• Rules can be constructed via a greedy search or a beam search. The latter may help to improve the quality of individual rules.
• Single-output, partial, or complete heads can be used by rules, i.e., they can predict for a single output, a subset of the available outputs, or all of them. Predicting for multiple outputs simultaneously enables to model local dependencies between them.
• Fine-grained control over the specificity/generality of rules is provided via hyperparameters.

Support for Post-Optimization and Pruning

• Incremental reduced error pruning can be used for removing overly specific conditions from rules and preventing overfitting.
• Post- and pre-pruning (a.k.a. early stopping) allows to determine the optimal number of rules to be included in an ensemble.
• Sequential post-optimization may help improving the predictive performance of a model by reconstructing each rule in the context of the other rules.

⌚ Runtime and Memory Optimizations

In addition to the features mentioned above, several techniques that may speed up training or reduce the memory footprint are currently implemented.

Approximation Techniques

• Unsupervised feature binning can be used to speed up the evaluation of a rule's potential conditions when dealing with numerical features.
• Sampling techniques and stratification methods can be used for learning new rules on a subset of the available training examples, features, or output variables.
• Gradient-based label binning (GBLB) can be used for assigning the labels included in a multi-label classification dataset to a limited number of bins. This may speed up training significantly when minimizing a non-decomposable loss function using rules with partial or complete heads.

Sparse Data Structures

• Sparse feature matrices can be used for training and prediction. This may speed up training significantly on some datasets.
• Sparse ground truth matrices can be used for training. This may reduce the memory footprint in case of large datasets.
• Sparse prediction matrices can be used for storing predicted labels. This may reduce the memory footprint in case of large datasets.
• Sparse matrices for storing gradients and Hessians can be used if supported by the loss function. This may speed up training significantly on datasets with many output variables.

Parallelization

• Multi-threading can be used for parallelizing the evaluation of a rule's potential refinements across several features, updating the gradients and Hessians of individual examples in parallel, or obtaining predictions for several examples in parallel.

Other Optimizations

• High- or low-precision data types can be used for gradients and Hessians to trade off training accuracy and speed differently.

📚 Documentation

Our documentation provides an extensive user guide, as well as Python and C++ API references for developers. If you are new to the project, you probably want to read about the following topics:

• Instructions for installing the software package or building the project from source.
• Examples of how to use the algorithm in your own Python code or how to use the command line API.
• An overview of available parameters.

A collection of benchmark datasets that are compatible with the algorithm are provided in a separate repository.

For an overview of changes and new features that have been included in past releases, please refer to the changelog.

📜 License

This project is open source software licensed under the terms of the MIT license. We welcome contributions to the project to enhance its functionality and make it more accessible to a broader audience. A frequently updated list of contributors is available here.

All contributions to the project and discussions on the issue tracker are expected to follow the code of conduct.