6 releases
0.0.6 | Apr 16, 2023 |
---|---|
0.0.5 | Apr 14, 2023 |
#517 in Machine learning
52 downloads per month
25KB
412 lines
H-Measure
The hmeasure crate provides a Rust implementation of the H-measure. The H-measure is a coherent alternative to the widely used AUC measure (Area Under the ROC Curve) for assessing the relative quality of binary classifiers. Whereas the AUC implicitly assumes a different cost distribution for the "cost of being wrong" depending on the classifier it is applied to, the H-Measure, by contrast, enables the researcher to fix the cost distribution consistently across all classifiers studied. The "cost of being wrong" is an inherent property of the subject being modelled and should not depend on the specific model being used to model the subject. H-Measure enables that consistency, whereas AUC does not.
The H-measure was introduced by David J. Hand in the paper:
"Measuring classifier performance: a coherent alternative to the area under the ROC curve" Mach Learn (2009) 77: 103–123 https://link.springer.com/article/10.1007/s10994-009-5119-5
A rust pyo3-based wrapper to expose the hmeasure crate in python is provided at: https://github.com/robinwiseman/pyhmeasure
Benchmark comparisons of the Rust hmeasure implementation with an equivalent python implementation are also provided.
A discussion with further examples is provided in Chapter 2 of: https://github.com/robinwiseman/finML/blob/aa12845f01454c24f36f4df0d1cb6e0993ea7c7f/src/finML_2022.pdf
Dependencies
~6.5MB
~123K SLoC