S. Benítez Peña, R. Blanquero, E. Carrizosa, P. Ramírez Cobo
Support Vector Machine (SVM) is a powerful tool to solve binary classification problems. However, SVM does not provide probabilities as other classifiers do in a natural way. Many attempts have been carried out in order to obtain those values, as in Sollich P. (2002) and Platt J. et al. (1999). Here, a bootstrap-based method yielding class probabilities and confidence intervals is proposed for a novel version of the SVM, namely, the cost-sensitive SVM in which misclassification costs are considered by incorporating performance constraints in the problem formulation. This is important in many contexts as creditscoring and fraud detection where misclassification costs may be different in different classes. In particular, our target is to seek the hyperplane with maximal margin yielding misclassification rates below given threshold values.
Palabras clave / Keywords: constrained classification, misclassification costs, mixed integer quadratic programming, sensitivity/specificity trade-off, Support Vector Machines, probabilistic outputs
Programado
Sesión GT02-4: Análisis Multivariante y Clasificación (AMyC-4). Organizadora: Eva Boj del Val
29 de mayo de 2018 17:00
Sala 5
