基于改进Adaboost软件缺陷数据特征选择方法

Feature selection method based on improved Adaboost for software defect prediction data

传统的Adaboost算法在处理软件缺陷数据时,面临两个问题:Adaboost未能考虑软件缺陷数据为不平衡数据,即无缺陷的样本数远远超过有缺陷的样本数;通过Adaboost选择出来的软件特征之间存在较大的相关性,这些特征会影响分类效果,影响预测结果。为此提出一种基于互信息及改进的Adaboost的集成算法MAboost。在NASA数据集上的实验结果表明,该算法对于软件缺陷数据具有较好的特征选择能力。

The traditional Adaboost algorithm of feature selection faces two problems in handling software defect prediction datasets,the first one is that imbalanced data namely samples of a class vastly outnumber the other class and the other is that high correlation among the selected features.In this situation,a feature selection method MAboost（mutual information and improved Adaboost based）was proposed to optimize the process.And the well-known NASA dataset was used for an empirical study to verify the competiveness of MAboost.