A1355
Title: CalCS: Calibrated cost-sensitive classification under strict error constraints
Authors: John Park - Hong Kong University (Hong Kong) [presenting]
Abstract: While fundamentally distinct in their objectives, the Neyman-Pearson (NP) and cost-sensitive (CS) paradigms both provide essential frameworks for classification. Standard CS methods often rely on empirical error estimates to satisfy user constraints. However, it is demonstrated that these plug-in estimators systematically underestimate true population risk. To resolve this, calibrated cost-sensitive classification (CalCS) is introduced, an algorithm that formally bridges the NP and CS paradigms. Given a strict NP error constraint, CalCS leverages exact finite-sample probability bounds to determine associated CS costs that provide the desired control. It is shown that the algorithm provides finite-sample guarantees while converging to the strict theoretical thresholds. Evaluations on synthetic and real-world datasets demonstrate that CalCS controls targeted error rates, successfully translating mathematical guarantees into robust, practical classifiers.
