%Aigaion2 BibTeX export from Idiap Publications %Saturday 23 November 2024 08:59:02 AM @TECHREPORT{Grandvalet_Idiap-RR-01-2009, author = {Grandvalet, Yves and Keshet, Joseph and Rakotomamonjy, Alain and Canu, St{\'{e}}phane}, projects = {Idiap, DIRAC}, month = {1}, title = {Support Vector Machines with a Reject Option}, type = {Idiap-RR}, number = {Idiap-RR-01-2009}, year = {2009}, institution = {Idiap}, abstract = {We consider the problem of binary classification where the classfier may abstain instead of classifying each observation. The Bayes decision rule for this setup, known as Chow{\^{a}}€™s rule, is de{\"{\i}}¬ned by two thresholds on posterior probabilities. From simple desiderata, namely the consistency and the sparsity of the classi{\"{\i}}¬er, we derive the double hinge loss function that focuses on estimating conditional probabilities only in the vicinity of the threshold points of the optimal decision rule. We show that, for suitable kernel machines, our approach is universally consistent. We cast the problem of minimizing the double hinge loss as a quadratic program akin to the standard SVM optimization problem and propose an active set method to solve it ef{\"{\i}}¬ciently. We {\"{\i}}¬nally provide preliminary experimental results We consider the problem of binary classi{\"{\i}}¬cation where the classi{\"{\i}}¬er may abstain instead of classifying each observation. The Bayes decision rule for this setup, known as Chow{\^{a}}€™s rule, is de{\"{\i}}¬ned by two thresholds on posterior probabilities. From simple desiderata, namely the consistency and the sparsity of the classi{\"{\i}}¬er, we derive the double hinge loss function that focuses on estimating conditional probabilities only in the vicinity of the threshold points of the optimal decision rule. We show that, for suitable kernel machines, our approach is universally consistent. We cast the problem of minimizing the double hinge loss as a quadratic program akin to the standard SVM optimization problem and propose an active set method to solve it ef{\"{\i}}¬ciently. We {\"{\i}}¬nally provide preliminary experimental results illustrating the interest of our constructive approach to devising loss functions.}, pdf = {https://publications.idiap.ch/attachments/reports/2008/Grandvalet_Idiap-RR-01-2009.pdf} }