What is Accuracy
A quality metric that shows what percentage of predictions the model made correctly.
Definition
Accuracy is often used in classification problems: for example, when a model determines whether an email is spam or a regular email. If out of 100 examples the model correctly recognized 90, its accuracy is 90%. But accuracy by itself is not always sufficient: for rare events, the model may look good, although it misses the most important cases.
Example
In the task of finding fraudulent transactions, a model may almost always say “everything is fine” and have a high percentage of correct answers, but still miss real attacks.
Why it matters
Understanding accuracy helps you not to believe one beautiful number and correctly compare AI tools, especially when the cost of an error is different for different classes.
How it works
The metric is calculated as the ratio of correct answers to the total number of verified examples. For an honest assessment, a separate test sample is used, which the model did not see during training.
Where it is used
- classifier evaluation
- comparison of models
- AI service quality control
Limitations
Precision is poorly suited for unbalanced data and problems where one error is more dangerous than another. Additional metrics are often needed: recall, positive class precision, F1 or AUC.