A0588
Title: Kick-one-out-based variable selection method for Euclidean distance-based classifier in high-dimensional settings
Authors: Tomoyuki Nakagawa - Meisei University (Japan) [presenting]
Hiroki Watanabe - Ferris University (Japan)
Masashi Hyodo - Kanagawa University (Japan)
Abstract: Classification is considered by using Euclidean distance-based classifier in high-dimensional data. The most important and difficult part of discriminant model building is selecting the appropriate variables for discriminant analysis. If the non-redundant variables that affect the classification rule are omitted, the expected probability of misclassification (EPMC) is large. On the other hand, EPMC may be large even if the redundant variables, which do not affect the classification rule, are included. Kick-One-Out(KOO)-based criteria are established for selecting non-redundant variables for the Euclidean distance-based classifier, and the consistency of the proposed variable selection in high dimensional settings is proved. The theoretical results are derived without assuming homogeneity of covariance matrices and multivariate normality for the group-conditional distribution.
