Fuzzy Twin Support Vector Machine Based on Smooth Pinball Loss

Twin support vector machine (TWSVM) is constructed based on hinge loss function and least squares loss function. And in TWSVM, all training samples are given the same importance. Thus, TWSVM is sensitive to noise, and is instability to re-sampling. To overcome the issue, a Fuzzy TWSVM with smooth Pinball loss (Sp-FTSVM) is proposed in this paper. By defining the smooth Pinball loss, Sp-FTSVM deals with the quantile distance, which reduces its sensitivity to noise. To further weaken the sensitivity to noise, different confidence levels are assigned to the samples by using fuzzy memberships. Lastly, the numerical experiments on original UCI datasets, corrupted UCI datasets with noise and steel surface defects dataset verify the feasibility and effectiveness of the Sp-FTSVM.