Boosting Kernel Discriminant Analysis for pattern classification
Shinji Kita, Seiichi Ozawa, Satoshi Maekawa, Shigeo Abe · 2007
This paper presents a new boosting algorithm called Boosting Kernel Discriminant Analysis (BKDA) in which the feature selection and the classifier training are conducted by Kernel Discriminant Analysis (KDA) and AdaBoost.M2, respectively. To reduce the dependency between classifier outputs and to speed up the learning, each classifier is trained in the different feature space which is obtained by applying KDA to a small set of hard-to-classify training samples. The proposed BKDA is evaluated using standard benchmark datasets. The experimental results demonstrate that BKDA outperforms both Boosting Linear Discriminant Analysis (BLDA) and Support Vector Machine (SVM) for multi-class classification problems. On the other hand, the performance evaluation for 2-class problems shows that the advantage of the proposed BKDA against BLDA and SVM depends on the datasets.