Do the concepts of residual and zero variance make sense when considering a classifier other than linear regression?

Question

Mathematics

Good afternoon. If I am considering a classifying neural network or a random forest, does it make sense to talk about metrics such as residual and zero deviation (residual deviance and null reviance)?

Answer 1

I will express my opinion. These indicators characterize not the method, but the result. That is, it doesn't matter what method you build your model with, their purpose is to evaluate how "good" the model is built.
On the other hand, these metrics have a semantic meaning in the case of solving problems of the regression type. For classification tasks - when the dependent feature is measured on a scale weaker than the interval one - the use of these indicators is meaningless. Criteria based on Pearson's contingency tables are already working there.
Thus, my answer to your question is - if you solve the regression problem using a neural network or a random forest - and this is quite possible - then using these metrics is acceptable. Otherwise, no.