Support vector machines for multi-class signal classification with unbalanced samples
Peng Xu, Andrew K. Chan · 2004
Support vector machines (SVMs) were originally developed for binary classification. To extend it to multi-class pattern recognition, one popular approach is to consider the problem as a collection of binary classification problems, so that each of them may be solved by a binary SVM. However, there is no guarantee that these SVMs will achieve the optimal solution even though each individual binary SVM is well trained. In this paper, we propose a method to optimize the multi-class SVMs by adjusting the penalty parameters using a genetic algorithm (GA). The method is applied to an acoustic signal classification problem with very promising results.