A Robust Fuzzy Support Vector Machine for Two-class Pattern Classification

Abstract

This paper proposes a systematic method to classify data with outliers. The essential techniques consist of the outlier detection and the fuzzy support vector machine (FSVM). In this approach, the main body set for each class is first determined by the outlier detection algorithm (ODA) that estimates the outliers based on the total similarity objective function. Then, incorporated with the total similarity measure of the ODA, a fuzzy membership degree is assigned to each training sample. Experiments show that the proposed method can greatly reduce the effects of outliers in the training process and the final decision surface of the FSVM is insensitive to outliers.