Corrected test statistics for comparing machine learning models on correlated samples
You can install the stable version of
You can install the development version of
Often in machine learning, we want to compare the performance of
different models to determine if one statistically outperforms another.
However, the methods used (e.g., data resampling, \(k\)-fold cross-validation) to obtain these
performance metrics (e.g., classification accuracy) violate the
assumptions of traditional statistical tests such as a \(t\)-test. The purpose of these methods is
to either aid generalisability of findings (i.e., through quantification
of error as they produce multiple values for each model instead of just
one) or to optimise model hyperparameters. This makes them invaluable,
but unusable with traditional tests, as Dietterich (1998)
found that the standard \(t\)-test
underestimates the variance, therefore driving a high Type I error.
correctR is a lightweight package that implements a small
number of corrected test statistics for cases when samples are not
independent (and therefore are correlated), such as in the case of
cross-validation, and repeated \(k\)-fold cross-validation. These
corrections were all originally proposed by Nadeau
and Bengio (2003). Currently, only cases where two models are to be
compared are supported.
A Python version of
correctipy is available at the GitHub