discriminability

Discriminability refers to the extent to which a system or measurement can distinguish between two or more stimulus categories, states, or distributions. It is a general notion used across psychology, statistics, signal processing, and machine learning to describe how separable signals are and how reliably differences can be detected.

In measurement and testing, discriminability describes how well a feature or test separates individuals or items into predefined groups. Factors include signal strength, noise, and the degree of overlap between the distributions of the groups. Quantitative summaries include classification accuracy, d prime, the area under the ROC curve (AUC), or the Fisher discriminant ratio. Higher discriminability corresponds to clearer separation.

In psychometrics, discriminability is linked to discriminant validity, the extent to which a measure is not strongly related to measures of different constructs. In test items, discrimination statistics describe how well an item differentiates between high- and low-performing examinees.

In neuroscience and pattern recognition, discriminability is used for multivariate pattern analysis and related methods: the ability of neural or feature patterns to distinguish experimental conditions. It is often assessed by cross-validated classifier performance or information-theoretic measures such as mutual information.

Limitations and considerations include the context-dependence of discriminability, potential bias from sampling or overfitting, and the trade-off between sensitivity and specificity. The term emphasizes separability rather than correctness alone, and its precise interpretation depends on the domain and chosen metrics.