Measuring accuracy, sensitivity, specificity, and precision for binary classification
When assessing a classification model, we typically want to know how often we are right. In the case of a binary target – one where the target has two possible categorical values – we calculate accuracy as the ratio of times we predict the correct classification against the total number of observations.
But, depending on the classification problem, accuracy may not be the most important performance measure. Perhaps we are willing to accept more false positives for a model that can identify more true positives, even if that means lower accuracy. This might be true for a model that would predict the likelihood of having breast cancer, a security breach, or structural damage in a bridge. In these cases, we may emphasize sensitivity (the propensity to identify positive cases) over accuracy.
On the other hand, we may want a model that could identify negative cases with high reliability...
