基于pinball损失的一对一加权孪生支持向量机

孪生支持向量机通过求解2个较小二次规划问题得到一对非平行超平面,从时间和准确率方面提高了分类器的性能.由于此方法使用Hinge损失函数,造成孪生支持向量机对噪声较为敏感以及重采样的不稳定.为此,针对多分类问题,将pinball损失函数与样本权重引入到孪生支持向量机中,采用一对一方法组合二分类器,提出了基于pinball损失的一对一加权孪生支持向量机,较好地解决了孪生支持向量机对噪声的敏感性以及重采样的不稳定性.另外,对于样本的不同影响,给出了多种求取样本权重的方法.实验中选取标准数据集和人工合成数据集对提出的算法进行了验证,并与一对一孪生支持向量机(OVO-TWSVM)、一对多孪生支持向量机(OVA-TWSVM)以及基于pinball损失的一对一加权孪生支持向量机(Pin-OVO-TWSVM)进行了比较,表明了提出方法的有效性.

One-versus-one weighted twin support vector machine with pinball loss function

Twin support vector machine obtains a pair of non-parallel hyperplanes by solving two smaller quadratic programming problems, which improves the performance of the classifier in terms of time and accuracy. Because this method uses the Hinge loss function, the twin support vector function is more sensitive to noise and the resampling is unstable. Therefore, in view of the multi-classification problem, the pinball loss function and weights of samples are introduced into the twin support vector machine, and the binary classifiers obtained are combined by a one-versus-one method,and a one-versus-one weighted twin support vector machine based on pinball loss is proposed. It solves the problems of the sensitivity of twin support vector machines to noise and the instability of resampling. In addition, for the different effects of samples, some methods for calculating the weights of samples are given. In the experiment, standard data sets and artificial synthetic data sets are selected to verify the proposed algorithm Pin-OVO-TWSVM and compared the performance with OVO-TWSVM, OVA-TWSVM and Pin-OVO-TWSVM to show the effectiveness of the proposed method.