Comparing Classifier's Performance Based on Confidence Interval of the ROC

Abstract

This paper proposes a new methodology for comparing} two performance methods based on confidence interval for the ROC curve. The methods performed and compared are two algorithms for face recognition. The novelty of the paper is three-fold: i) designing a methodology for the comparison of decision making algorithms via confidence intervals of ROC curves; ii) investigating how sample sizes influence the properties of the particular methods; iii) recommendations for a general comparison of decision making algorithms via confidence intervals of ROC curves. To support our conclusions we investigate and demonstrate several approaches for constructing parametric confidence intervals on real data. Thus, we present a non-traditional and reliable way of reporting pattern recognition results using ROC curves with confidence intervals.