基于特征缺省的最小类内方差支持向量机

最近提出的基于特征缺失的支持向量机(support vector machine with absent features,AF-SVM)在处理具有特征缺失的数据分类时,得到的分类超平面不能很好地适应数据的总体分布,并存在两类误分的比例相差比较大的问题。为此,本文通过引入最小类内方差支持向量机(minimum class variance SVM, MCVSVM)分类机制,提出了基于特征缺失的最小类内方差支持向量机(minimum within-class variance SVM with absent features,AF-V-SVM)。AF-V-SVM一方面可以依据数据集的分布特性,改善分类超平面的方向性;另一方面,通过自由设置分类间隔的定义空间,调整误分的比例。实验表明,与其他基于特征缺省的分类方法相比,该方法不仅提高了分类正确率而且使分类效果更加合理。

Minimum within-class variance SVM with absent features

In the classification of data with absent features, the lately proposed support vector machine with absent features (AF SVM) has some drawbacks: the obtained classification hyper plane with AF-SVM can not well adapt to the data’s overall distribution, and the proportion of the misclassified data differs greatly between the two classes. To overcome these drawbacks, a minimum within-class variance SVM with absent features (AF-V-SVM) was proposed based on the technology of minimum class variance SVM (MCVSVM).On the one hand, AF-V-SVM could improve the direction of the classification hyper plane with the information of the distribution feature of the data set; on the other hand, this method adjusted the proportion of misclassified data by freely setting the definition space of classification margin. Experiments showed that the method in this paper was superior to other absent features based classification methods in the aspects of classification accuracy and rationality.