Multiple support vector domain classifier

作者：Wu De; Liu Sanyang^*

来源：Journal of Xi'an Jiaotong University, 2012, 46(6): 87-91.

• Class-distance
• Minimal enclosing ball
• Multiple classifiers
• Space position
• Support vector domain description

To solve the difficulties in multi SVM, such as huge computation and long training time, a multiple support vector domain classifier (MSVDC) is proposed. In the training process, the support vector domain description (SVDD) is employed to obtain the minimal enclosing ball (MEB) of each class, and then the data space is divided into different regions. In the test phase, the distances from the test sample to the MEB centers are evaluated, and the position of the test sample is determined. For samples in the overlapped and outside regions of the MEB, a relative class distance is defined, and is erected in the class with the smallest value. MSVDC avoids the repeated usage of training data, and reduces the memory and enhance the efficiency. Numerical experiments show that MSVDC is endowed with better robustness. The classification accuracy gets to 98.89%, 4.51% and 1.24% higher than "one-against-all" and "one-against-one", and the training time is only spent for 18.06% and 55.41% of "one-against-all" and "one-against-one", respectively.

• 出版日期2012
• 单位西安电子科技大学