⚖ Weighting Schemes Report

Extend the scikit-learn classification report with custom weighting schemes.

🧰 Usage

from weightings_schemes_report import classification_report
classification_report(
    y_true,
    y_pred,
    average_funcs=('entropy', 'dodrans'),
    *args,
    **kwargs)

Package for the paper "Why only Micro-F1? Class Weighting of Measures for Relation Classification". We provide an extended classification report that gives more F1 scores than only micro, macro and weighted. Custom additional weighting schemes can be added to average_funcs by defining in custom_scoring and giving the string of the function name to average_funcs. A function added to custom_scoring should take an 1d numpy array of class counts and a numpy array of labels and return (not necessarily normalized) class_weights.

Functionality of classification_report should otherwise be identical to the classification_report from scikit-learn.

📚 Citation

If you find the code helpful, please consider citing the following paper:

@inproceedings{harbecke2022why,
    title={Why only Micro-F1? Class Weighting of Measures for Relation Classification},
    author={David Harbecke and Yuxuan Chen and Leonhard Hennig and Christoph Alt},
    year={2022},
    booktitle={Proceedings of NLP Power! The First Workshop on Efficient Benchmarking in NLP}
}