具有间隔分布的抗噪声非平行支持向量机

Anti-Noise Nonparallel Support Vector Machine with Margin Distribution

针对非平行支持向量机(NonParallel Support Vector Machine,NPSVM)对噪声敏感和忽略了数据分布结构的问题,提出了一种具有间隔分布的抗噪声非平行支持向量机(Anti-Noise NPSVM with Margin Distribution, MDANPSVM)分类模型.在MD-ANPSVM模型中,每个优化问题同时最小化两类样本的基于L1范数的绝对损失和改进的铰链损失,这可以保证模型的稳定性,减小噪声和异常值的影响.此外,在MD-ANPSVM模型中,采用一阶和二阶统计量来描述训练数据的间隔分布信息,并试图同时最大化间隔均值和最小化间隔方差,这进一步提高了模型的泛化性能.最终,我们在不同的数据集上进行了对比实验.实验结果显示,MD-ANPSVM模型具有较强的泛化能力和强鲁棒性.

Because nonparallel support vector machine(NPSVM)is sensitive to noise and ignores the distributing structure of data,an anti-noise NPSVM with margin distribution(MD-ANPSVM)model is proposed.In MD-ANPSVM,each optimization problem simultaneously minimizes the L1-norm loss and improved hinge loss,which can ensure the sta⁃bility of the model and reduce the adverse impact of noise and outliers.In addition,in MD-ANPSVM,the margin distribu⁃tion described by the first-and second-order statistics is introduced.Each optimization problem simultaneously maximizes the margin mean and minimizes the margin variance,which results in better generalization performance.The experimental results on the UCI datasets and steel surface defects dataset show that MD-ANPSVM can achieve better generalization abili⁃ty and strong robustness.