Evaluation of Feature Selection by Multiclass Kernel Discriminant Analysis

摘要

In this paper, we propose and evaluate the feature selection criterion based on kernel discriminant analysis (KDA) for multiclass problems, which finds the number of classes minus one eigenvectors. The selection criterion is the sum of the objective function of KDA, namely the sum of eigenvalues associated with the eigenvectors. In addition to the KDA criterion, we propose a new selection criterion that replaces the between-class scatter in KDA with the sum of square distances between all pairs of classes. To speed up backward feature selection, we introduce block deletion, which deletes many features at a timeC and to enhance generalization ability of the selected features we use cross-validation as a stopping condition. By computer experiments using benchmark datasets, we show that the KDA criterion has performance comparable with that of the selection criterion based on the SVM-based recognition rate with cross-validation and can reduce computational cost. We also show that the KDA criterion can terminate feature selection stably using cross-validation as a stopping condition.