How to Prevent Model Accuracy from Tricking You
At least mathematically speaking, the criteria used to evaluate the performance of classification models are rather simple. But I've noticed that many modelers and data scientists have trouble expressing these measures, and some even use them inappropriately. This is a simple error to make because these metrics seem straightforward at first glance, but depending on the issue domain, they may have significant ramifications.
Ninety subjects—representing whatever we could want to classify—are included in each visualization. Red subjects are positive samples, whereas blue subjects are negative samples. The model that tries to forecast positive samples is shown in the purple box. The model predicts anything inside this box to be good.
This article acts as a picture manual.
0 Comments