search

Arn-BAuA / TimeSeriesAEBenchmarkSuite

This Repository contiains a benchmark environment to compare various autoencoders on multiple time series datasets.
GNU General Public License v3.0
2 stars 0 forks source link

Run time optimisation with AUC-Score #2

Open Arn-BAuA opened 12 months ago

Arn-BAuA commented 12 months ago

The method for the AUC-Score calculation runs very poorly. It would be great to implement an adaptive quadrature in C++ and include it as a submodule to substitute the current version of the calculation.

Arn-BAuA commented 11 months ago

Update: Upon researching the AUC-Score I found out, that the AUC-Score is identical to the Wilcoxon–Mann–Whitney test (Hanley, James A and McNeil, Barbara J: The meaning and use of the area under a receiver operating characteristic (ROC) curve. DOI: https://doi.org/10.1148/radiology.143.1.7063747). Computing the AUC-Score via quadrature from the ROC is in deed a bad idea. Not only does it introduce errors from the numeric integration, it also completely ignores the statistic errors that emerge from the sampling of the underlying distributions of normal and abnormal data. So instead of a quadrature error, like we are reporting at the moment, the correct way of reporting the error would be to use a statistic estimate for the variance like the one proposed by Qing Wang, Alexandria Guo (Qing Wang, Alexandria Guo: An efficient variance estimator of AUC and its applications to binary classification: DOI: 10.1002/sim.8725). We will create a python version based on their work in the following days to evaluate AUC and the uncertainty for the AUC.