Set-Valued Support Vector Machine with Bounded Error Rates

This article concerns cautious classification models that are allowed to predict a set of class labels or reject make prediction when the uncertainty in is high. set-valued approach equivalent task acceptance region learning, which aims identify subsets input space, each guarantees cover observations with at least predetermined probability. We propose directly learn regions through risk minimization, by making use truncated hinge loss and constrained optimization framework. Collectively our theoretical analyses show these regions, high probability, satisfy simultaneously two properties: (a) they guarantee noncoverage rate bounded from above; (b) give ambiguous predictions among all satisfying (a). An efficient algorithm developed numerical studies conducted using both simulated real data. Supplementary materials for this available online.