Abstract:

The traditional support vector machine (SVM) often falls into overfitting when outliers are contained in the training data. The fuzzy support vector machine can effectively deal with this problem. According to the deficiency of the membership function designed based on the distance between a sample and its cluster center, a novel fuzzy support vector machine based on the class centripetal degree (CCDFSVM) is proposed. It combines the distance between a sample and its cluster center with the relationship between samples expressed as the class centripetal degree. This function can effectively separate the valid samples from the noises or outliers. Besides, the size of the class centripetal degree can reflect the samples mixed degree. Experimental results show that the fuzzy support vector machine based on the class centripetal degree is more robust than the traditional support vector machine, and it outperforms the other two FSVM counterparts with different membership functions in terms of antinoise and classification performance.

Key words: fuzzy support vector machine;membership function;class centripetal degree