非均衡数据情形的一种协同正则化多视图半监督学习分类器

A multi-view based semi-supervised classifier with co-regularization for imbalanced data

利用多视图学习、流形学习以及协同正则化的多重惩罚处理,对含有大量无标签的类别数据提出一种多视图半监督学习的分类器构造方法.该方法由递归提升的方式对数据进行逐次多视图半监督学习,通过适当的标签化、均衡化处理改进每次集成的学习效率直到稳定.通过最小二乘和多分类SVM研究了新方法的性质,给出泛化误差的一个有意义上界,体现了新方法良好的泛化能力.模拟研究和实证分析显示,在有限样本情形下新方法具有良好的表现.

A method of constructing a multi-view semi-supervised learning classifier was presented for manifold learning and multi-puncture processing.The multi-view and semi-supervised learning of the data is achieved through recursive optimization,and appropriate labeling and equalization processing,until the efficiency of learning becomes stable.The properties of this multi-classifier were given,for instance,an upper bound of the generalization error,which showed a good capacity for generalization.Simulation and empirical analysis showed that the new method performs well with small samples.