Accuracy

Properties2

Type	Concept
Note created	Feb 17, 2025

Accuracy is a classification error metric that measures how close or far off a given set of predictions are of their expected labels. Its most widespread definition is a description of systematic errors; a measurement of statistical bias. Low accuracy causes a difference between a prediction and a true value. It is normally coupled with (and even mistaken for) precision.

The formal expression of it is:

$$\text{Accuracy}=\frac{\text{correct classifications}}{\text{all samples}}$$

In binary classification, the correct classifications are the sum of both true positives and true negatives. In multiclass classification this is also referred to as top-1 accuracy, to distinguish it from other less-unforgiving metrics like top-5 accuracy, where a prediction is correct if the correct class falls within the top 5 predicted labels.