Efficient Binary Tree Multiclass SVM using Genetic Algorithms for Vowels Recognition

摘要

In this paper we introduce and investigate the performance of a simple framework for multiclass problems of support vector machine (SVM), we present a new architecture named EBTSVM (Efficient Binary Tree Multiclass SVM), in order to achieve high classification efficiency for multiclass problems. The proposed paradigm builds a binary tree for multiclass SVM by genetic algorithms with the aim of obtaining optimal partitions for the optimal tree. Our approach is more accurate in the construction of the tree. Further, in the test phase EBTSVM, due to its Log complexity, it is much faster than other methods in problems that have big class number. In the context of phonetic classification by EBTSVM machine, a recognition rate of 57.54%, on the 20 vowels of TIMIT corpus was achieved. These results are comparable with the state of the arts, in particular the results obtained by SVM with one-versus-one strategy. In addition, training time and number of support vectors, which determine the duration of the tests, are also reduced compared to other methods. However, these results are unacceptably large for the speech recognition task. This calls for the development of more efficient multi-class kernel methods in terms of accuracy and sparsity.