Quantitative Measurements of model interpretability for the analysis of spectral data
Andreas Backhaus, Udo Seiffert · 2013
Classically, machine learning methods are evaluated according to their accuracy and model size. Increasingly model parameters are used to interpret the model in order to extract information about the data it was build on. The capability of a model to deliver this kind of information, its interpretability, is so far more or less subjective. In this paper a number of quantitative measures are suggested to compare machine learning methods in their capability to offer interpretation of the underlying data.