R. Blanquero, E. Carrizosa, C. Molero-Río, D. Romero Morales
Random Forests are a powerful prediction tool obtained by bagging decision trees. Classical decision trees are defined by a set of orthogonal cuts, i.e., the branching rules are of the form variable X not lower than threshold c. The variables and thresholds are obtained greedily. The use of a greedy strategy yields low computational cost, but may lead to myopic decisions. Although oblique cuts, with at least two variables, have also been proposed, they involve cumbersome algorithms to find each cut of the tree. The latest advances in Optimization techniques have motivated further research on new procedures to build optimal classification trees. In this talk, we propose a non-linear continuous programming formulation to tackle this issue. Our numerical results show the usefulness of this approach: using one single tree, we obtain better performance than classification trees and close to Random Forests, being much more flexible since class performance constraints can be easily included.
Palabras clave / Keywords: classification, decision trees, non-Linear continuous optimization
Programado
Sesión GT02-1: Análisis Multivariante y Clasificación (AMyC-1). Organizadora: Eva Boj del Val
29 de mayo de 2018 10:30
Sala 5
