Item

Abstract

As a variant of Support Vector Machine (SVM), Large Margin Distribution Machine (LDM) has been validated to outperform SVM both theoretically and experimentally. Due to the inevitable noise in real applications, the credibility of different samples is not necessarily the same, which is neglected by most existing LDM models. To tackle the above problem, this paper first introduces fuzzy set theory into LDM, and proposes a Fuzzy Large Margin Distribution Machine (FLDM) with better robustness and performance. Considering the noise and uncertainty in datasets, sample points farther from the center of homogenous class are less reliable. Therefore, a fuzzy membership function based on the distance to the class center is utilized to characterize the confidence of each sample, i.e., the degree to which the sample belongs to a certain category. Furthermore, different strategies are developed to obtain class centers for linearly separable and linearly inseparable problems. Experiments conducted on both artificial and UCI datasets verified the superiority of FLDM from different perspectives.