##
**Regularization through fuzzy discrete SVM with applications to customer ranking.**
*(English)*
Zbl 1269.68098

Summary: In the context of classification most efforts have been devoted to deriving accurate prediction models from a set of examples whose class is supposed to be known with certainty. However, there are situations where class labels are affected by an intrinsic vagueness, as in ranking customers for marketing campaigns or credit approval. In this paper we propose a new two-phase fuzzy classification method aimed at generating accurate classification rules when labels are uncertain. In the first phase, an ensemble method is applied in order to derive the value of the class membership function for each example in the dataset. In the second phase, an optimal classification model is obtained by solving a fuzzy variant of discrete support vector machines. Computational tests performed on benchmark and real world marketing and credit risk datasets show the effectiveness of the proposed method when it is compared to alternative classification techniques. Furthermore, the tests reveal that the new fuzzy discrete SVM model is a robust regularization method capable of generating stable classification rules, reducing the variance of the error and smoothing out the noise due to outliers.