A0555
Title: Confidence interval for recall and precision of multi-class classification
Authors: Kanae Takahashi - Hyogo Medical University (Japan) [presenting]
Kouji Yamamoto - Yokohama City University (Japan)
Abstract: In the medical field, binary classification problems are common, and accuracy, sensitivity, specificity, negative predictive value, and positive predictive value are often used as indicators of the performance of binary predictors. Also, in computer science, classifiers are usually evaluated with recall (sensitivity) and precision (positive predictive value). Recall and precision are only applicable to binary classification data. Two aggregate performance measures have been proposed for recall and precision in multi-class classification problems: macro-averaged recall and precision (maR, maP) and micro-averaged recall and precision (miR, miP). The maR is the arithmetic mean of recall for each class and maP is the arithmetic mean of precision for each class. On the other hand, miR and miP are the recall and precision computed from the sum of the decisions per sample. Most articles report point estimates of recall and precision for multi-class classification without considering the uncertainty of the estimates. Therefore, we propose methods for estimating maR, maP, miR, and miP with confidence intervals based on the large sample multivariate central limit theorem and the delta method.
