【摘要】 针对标准数据集在评估多分类器系统的组合方法时存在的不足,设计了一种新的分类器模拟算法.该算法利用分类器的识别率建立混淆矩阵,由混淆矩阵生成基分类器的决策,进而结合分类器之间的相关性度量生成所有的模拟数据.通过实验评估表明,该算法能够模拟任意多个分类器和任意多个模式类别的数据,且能够表达出分类器之间的关联性.又应用生成的模拟数据集对多数投票和堆叠泛化这2种组合方法进行了实验,结果表明分类器之间的负相关有助于提高系统的性能,特别是当单个分类器识别率取0.8、关联度从0.829 5降至-0.484 7时,多数投票和堆叠泛化的性能分别提高了14.98%和41.99%.更多还原

【Abstract】 Aiming at the deficiency of evaluating classifier combination methods with standard data sets,a new classifier simulation algorithm was proposed.The confusion matrix was established by the classifier’s recognition rate,by which the decision of the base classifier was made.Then all simulating data were generated by combining the correlation measure between classifiers.Through experimental investigation it is shown that the algorithm can simulate any number of classifier and any number of data with any kind of pattern,and it can also express the dependency between classifiers.With simulated datasets,experiments were carried out on both of majority vote and stacking combination method.The results indicate that negative correlation can improve the classification performance.The accuracy of the two methods increases by 14.98% and 41.99% respectively when decreasing the correlation from 0.829 5 to-0.484 7,particularly when the recognition rate of individual classifier is set to 0.8.更多还原