Identification of uncertainty and decision boundary for SVM classification training using belief function

摘要

The existence of noisy samples increases the inefficiency of SVM training. In SVM training, the classification hyperplane is determined by the support vectors, therefore, the other samples do not affect the SVM classifier. This paper presents a novel method that does not use all samples for SVM training. The basic idea is a novel method in which, at the first step, using the belief function theory and fuzzy rough set theory, the boundary samples are identified. And at the second step, the boundary samples uncertainty such as noisy samples are identified and discarded. Finally, at the last step, using the obtained boundary samples, the training of the SVM classifier is done. To show the performance of the proposed method, BFFR-BS (Belief Function and Fuzzy Rough Set-Boundary Samples), several experiments have been conducted on various real-world data sets from UCI repository. Experimental results reveal that the proposed method is superior to the state-of-the-art competing methods regarding accuracy, precision, time, and AUC metrics.