On Support Vector Machines Under a Multiple-Cost Scenario

S. Benítez-Peña, R. Blanquero, E. Carrizosa, P. Ramírez-Cobo

Support Vector Machines (SVM) is a benchmark procedure in Supervised Classification.
The classifier is obtained by solving a nonlinear convex optimization problem, in which a distribution-free proxy of the probability of misclassification is minimized.
While the misclassification cost may be an appropriate performance measure in some cases, there are many problems of practical interest in which different measures should better be taken into account; these include the misclassification rates in different classes, the predictive values error for the different classes, as well as measurement costs (an issue related but more complex than traditional feature selection).

In this work we propose an algorithmic approach to address the problem of building the SVM classifier, and also tuning the kernel parameters, by taking into account surrogates of all such costs.

Palabras clave: Support Vector Machines, Costs, Mixed Integer Nonlinear Programming

L05.1 Grupo de Análisis Multivariante y Clasificación I
5 de septiembre de 2016  11:30
0.02 - Aula de proyectos 1

E. Carrizosa, V. Guerrero, D. Romero Morales

C. Araya Alpizar

M. Agulló Antolín, E. Del Barrio Tellado