The proportion of a system's outputs or predictions that exactly match the true or intended values; in computational evaluation, usually reported as the fraction of correct results over all cases.

Principle

Measure overall correctness by counting true matches against a known ground truth; treats all errors equally and aggregates across classes or outcomes.

Demonstration

In a binary classifier tested on 100 labeled examples where 82 predictions equal their labels, accuracy = 82/100 = 0.82.

Misapplication

Using accuracy on a heavily imbalanced dataset to claim good performance when a trivial classifier that always predicts the majority class attains high accuracy while failing on the minority class.

Consequence

When used appropriately with balanced or appropriately weighted datasets, accuracy provides a single-number summary of correctness that is easy to interpret and compare.

Reversal

Focusing on per-class recall or precision instead of overall accuracy highlights class-specific performance and can reveal poor behavior hidden by a high accuracy number.

Boundary

Applies to problems with a well-defined ground truth and discrete evaluable outputs; excludes continuous error metrics (e.g., mean squared error) and contexts where class imbalance makes raw proportion misleading without weighting or complementary metrics.

Semantic Tension

Often confounded with 'precision' and 'recall' in information-retrieval contexts; accuracy is an aggregate correctness measure, whereas precision measures positive predictive value and recall measures true-positive rate.

Synthesis

Accuracy is the aggregate fraction of correct outputs relative to a defined ground truth; it is simple and informative for balanced, discrete tasks but must be complemented by class-aware or continuous metrics when distributional or severity differences matter.