基于选择性的贝叶斯分类器集成方法

提出一种基于选择性的贝叶斯分类器集成方法．该方法为避免数据预处理时的属性约简对分类效果的直接影响,在训练集上通过随机属性选取生成若干属性子集,并以这些子集训练不同的贝叶斯分类器,进而采用遗传算法优选贝叶斯分类器集成,其适应度函数综合了分类器的精度和差异度两项指标．实验中,将该方法与已有方法在UCI的标准数据集上进行了性能比较,并将该方法用于C^3I系统中的威胁度估计。

A Selective Approach for an Ensemble of Simple Bayesian Classifiers

To avoid the influence of feature reduction from data pre-processing on the performance of classification, a selective approach was proposed for an ensemble of simple Bayesian classifiers(ESBC), making use of random feature selection to generate several feature subsets from the whole training set, and obtained different simple Bayesian classifiers(SBCs) with the feature subsets, and then optimized the ESBC using genetic algorithms (GA), wherein the fitness function of GA involved the accuracy and diversity of SBCs. In the experiments, this approach was compared with existing methods in their performance through some standard data sets from the UCI machine learning repository, and was applied to the threat-degree estimation in C 3I.