【摘要】 针对分类器的构建,在保证基分类器准确率和差异度的基础上,提出了采用差异性度量特征选择的多分类器融合算法(multi-classifier fusion algorithm based on diversity measure for feature selection,MFA-DMFS)。该算法的基本思想是在原始特征集中采用Relief特征评估结果按权值大小选择特征,构造特征子集,通过精调使各特征子集间满足一定的差异性,从而构建最优的基分类器。MFA-DMFS不但能提高基分类器的准确率,而且保持基分类器间的差异,克服差异性和平均准确率之间存在的相互制约,并实现这两方面的平衡。在UCI数据集上与基于Bagging、Boosting算法的多分类器融合系统进行了对比实验,实验结果表明,该算法在准确率和运行速度方面优于Bagging和Boosting算法,此外在图像数据集上的检索实验也取得了较好的分类效果。更多还原

【Abstract】 The fusion of multiple classifiers is an important means of improving the efficiency of pattern recognition.The text,from the viewpoint of the accuracy of the component classifier as well as weight differences between the two classifiers,proposed that the multi-classifier fusion algorithm,based on differences in measurement,fell into a two-stage classifiers used to build the characteristics of the different components of a subset,the relief characteristics of the assessment results in accordance with the weights,were firstly used,from the original feature set to choose features to make into feature subset selection,and then by the fine-tuning to enable the feature subset to meet certain differences.Compared with the result of the experiments,it was found that the accuracy of the algorithm is always better than Bagging and the multiple classifiers fusion system constructed by Boosting algorithm.Moreover,its running speed is much higher than Bagging and Boosting algorithms.Finally,it was found that,in image database retrieval experiments,it has achieved better classification results.更多还原