Accuracy
In one sentence Accuracy is the fraction of all predictions the model got right, which becomes misleading when one class is rare.
Updated
Accuracy is the share of all predictions that were correct, counting every class together.
It is the exam-score metric: 87 questions right out of 100 is 87% accuracy. Nothing about it is subtle, and that is both its charm and its trap.
The trap is imbalanced data. Suppose 1 in 1,000 card payments is fraud. A "model" that predicts every payment is honest scores 99.9% accuracy — while catching zero fraud. The headline number looks superb and the model is worthless.
1,000 payments: 999 honest, 1 fraud
model says "all honest" → 999 correct → 99.9% accuracy, 0 fraud caughtWhen to trust it
Accuracy is a fair summary when classes are roughly balanced and all mistakes cost about the same — recognising handwritten digits, for example. The moment one class is rare, or one mistake is costlier than another, switch to precision, recall, or the full confusion-matrix, which show where the errors actually land.
Where to go next
- Full lesson: Model evaluation
- Related terms: precision, recall, class-imbalance, confusion-matrix