Abstract:

Support vector machine is one of the most popular binary classification algorithms,but data sets in the real world require multi-classification. Directed Acycline Graph (DAG) is one of the most used ways that expand SVM to support multiclass classification.DAG calls the classifiers less frequently and works faster than other methods.However,the accumulated mistakes cannot be cleared, and it has k! kinds of decision structures when dealing with k-class problems.Therefore structure selection becomes a key problem while using DAG-SVMs.In this paper we propose a highly accurate DAG structure selection method that uses the classificatory percentage in the training data sets to estimate the accuracy of the test data sets, and chooses the DAG structure with the highest accuracy. Experimental results show that compared with other methods,the proposed method can improve the classification accuracy of test data set dramatically and has a better effect in performing multi-class classification of the data sets without too many different types.

Key words: support vector machine;multi-classification;DAGSVM;structure selection