accuracy
/ AK-yur-uh-see /
Accuracy is the simplest report card a classifier can get: out of every prediction it made, what fraction did it get right? Mark 100 emails as spam-or-not, get 92 of them correct, and your accuracy is 92 percent. It is the number most people reach for first because it answers the obvious question — how often is the thing right?
Formally, accuracy is the count of correct predictions divided by the total number of predictions, lumping together every kind of correct answer (both correct yeses and correct noes). It runs from 0 to 1, or 0 to 100 percent, and bigger is better. It needs labeled examples whose true answer you already know, so you can compare the model's guess against reality.
Its honest weakness is that it can be wildly misleading when the classes are lopsided. Suppose 99 of every 100 transactions are legitimate and only 1 is fraud. A lazy model that simply declares everything legitimate scores 99 percent accuracy while catching exactly zero fraud — useless, yet it looks excellent. Whenever one outcome is rare, accuracy alone hides the failure that matters, and you must look at precision, recall, or the confusion matrix instead.
A medical screen tests 1,000 people; 10 truly have the disease. The model labels everyone "healthy." It is wrong on only 10 people, so accuracy = 990/1000 = 99%. Yet it caught none of the sick — a perfect example of why 99% accuracy can mean a worthless model.
High accuracy, zero usefulness: why imbalanced data breaks the simplest metric.
Always ask the class balance before trusting an accuracy number. On a 50/50 split it is meaningful; on a 99/1 split it is nearly worthless on its own. A useful sanity check: how does your model compare to the trivial "always predict the majority class" baseline?