KI-Glossar

True Positive Rate

Proportion of positive samples correctly identified by the model, also known as sensitivity or recall.

False Positive Rate

Proportion of negative samples incorrectly classified as positive, corresponding to 1 minus specificity.

Decision Threshold

Threshold value that converts the output probabilities of a classifier into discrete binary predictions.

Optimal Operating Point

Point on the ROC curve that maximizes a composite metric like Youden's index or minimizes a specific cost function.

Precision-Recall Curve

Alternative graph to the ROC curve visualizing the trade-off between precision and recall, particularly useful for imbalanced classes.

Youden's Index

Evaluation metric combining sensitivity and specificity, defined as their sum minus 1, used to select the optimal threshold.

DET Curve

Detection Error Tradeoff graph using logarithmic scales to more clearly visualize classifier performance in low error rate regions.

PR Curve

Abbreviation for Precision-Recall curve, an alternative to the ROC curve preferred when the cost of false positives differs from that of false negatives.

Lift curve

Graph evaluating the performance improvement of a model compared to a random model, measuring the targeted gain ratio.

Brier score

Calibration metric evaluating the accuracy of predicted probabilities, complementary to AUC which only measures discrimination.

DeLong test

Non-parametric statistical test allowing for significant comparison of AUCs of two or more classifiers on the same dataset.

Multiclass ROC curve