Scikit Multilearn Versions Save

A new feature release:

• first python implementation of multi-label SVM (MLTSVM)
• a general multi-label embedding framework with several embedders supported (LNEMLC, CLEMS)
• balanced k-means clusterer from HOMER implemented
• wrapper for Keras model use in scikit-multilearn

Fix a lot of bugs and generally improve stability, cross-platform functionality standard and unit test coverage. This release has been tested with a large set of unit tests that work across Windows

Also, new features:

• multi-label stratification algorithm and stratification quality measures
• a robust reorganization of label space division, alongside with a working stochastic blockmodel approach and new underlying layer - graph builders that allow using graph models for dividing the label space based not just on label co-occurence but on any kind of network relationships between labels you can come up with
• meka wrapper works fully cross-platform now, including windows 10
• multi-label data set downloading and load/save functionality brought in, like sklearn's dataset
• kNN models support sparse input
• MLARAM models support sparse input
• BSD-compatible label space partitioning via NetworkX
• dependence on GPL libraries made optional
• working predict_proba added for label space partitioning methods
• MLARAM moved to from neurofuzzy to adapt
• test coverage increased to 94%
• Classifier Chains allow specifying the chain order
• lots of documentation updates

• a general matrix-based label space clusterer has been added which can cluster the output space using any scikit-learn compatible clusterer (incl. k-means) support for more single-class and multi-class classifiers you can now use problem transformation approaches with - your favourite neural networks/deep learning libraries: theano, tensorflow, keras, scikit-neuralnetworks support for label powerset based stratified kfold added
• graph-tool clusterer supports weighted graphs again and includes stochastic blockmodel calibration
• bugs were fixed in: classifier chains and hierarchical neuro fuzzy clasifiers

• *kNN classifiers support sparse matrices properly
• support for the new model_selection API from scikit-learn
• extended graph-based label space clusteres to allow taking probability of a label occuring alone into consideration
• compatible with newest graphtool
• support the case when meka decides that an observation doesn't have any labels assigned
• HARAM classifier provided by Fernando Benitez from University of Konstanz
• predict_proba added to problem transformation classifiers
• ported to python 3