基于Universum的多视角全局和局部结构风险最小化模型

Universum-based multi-view global and local structural risk minimization model

为克服传统多视角分类器无法充分最小化结构风险的不足,提出基于Universum的多视角全局和局部结构风险最小化模型。该模型采用Universum学习,利用有标签样本生成大量包含分类信息的无标签样本,从而增加分类器性能。这些信息有利于最小化结构风险。通过在Mfeat、Reuters和Corel等3个多视角数据集上的试验可以发现,该模型可以提高多视角分类器的性能,并可以更好地应用到多视角数据集的分类问题中。

In order to overcome the disadvantage of traditional multi-view classifiers that can not fully minimize structural risk,a Universum-based multi-view global and local structural risk minimization model is proposed.The model uses Universum learning,which uses labeled samples to generate a large number of unlabeled samples containing classification information so as to enhance the performances of classifiers.This information helps minimize structural risks.Experiments on three multi-view data sets,i.e.,Mfeat,Reuters and Corel,show that the model can improve the performance of multi-view classifiers and can be better applied to the classification of multi-view data sets.