What is accuracy? The fraction of all predictions that are correct
Explain classification accuracy, distinguish precision and recall and assess imbalance and critical errors.
Understand the concept
Classification accuracy is the fraction of all correct predictions, including positive and negative cases. Precision instead measures correctness among predicted positives. Accuracy needs references and scope; high values can hide missed rare classes. Generation also needs explicit correctness criteria rather than counting valid format as factual success.
The denominator is all samples, not one label
Google and scikit-learn distinguish these metrics. Accuracy uses all samples and can look high under class imbalance: always predicting ordinary mail may miss every spam message. Interpret errors and rare classes with the aggregate. Reports should specify mathematical accuracy rather than an undefined everyday judgment.
Use fixed references to inspect each decision
Compare predictions against human-confirmed public labels, then calculate overall correctness. Include rare classes and predefined missing-input behavior. Do not revise answers after results. No classifier or dataset score was run here. Open answers need factual and task criteria rather than a keyword-presence test for whole-answer correctness.
Data and rule changes alter score meaning
Version samples and labels and document multiclass or multilabel implementations. Include reference and review effort. Report precision and recall separately when needed. A score supports the evaluated distribution, not the same probability for every future answer; aggregate accuracy is not per-response confidence.
Checked sources
These sources support the concepts and methods described here. Examples and test scope are labeled separately; check the version of the tools you use.