Generalization Ability in SVM with Fuzzy Class Labels

The paper attempts to introduce a fundamental fuzzy concept to break the equivalent attitude of the input training set of SVM, and tries to give individual example in the set a different attitude. The attitude can stand for the influence that the example takes into account in the classification. In the paper, we present a method to refresh the attitude by assigning proper fuzzy value to the class label of each example. Based on the benefit of individualized fuzzification, we re-examine the formulation of SVM, and try to find out the corresponding effects. Although it costs a little more computation, the introduction of fuzzy class label is eventually worth for the better generalization performance with a large margin. The effect is especially both crucial and subtle in worse confused training data set with many hard examples