面向软件缺陷个数预测的混合式特征选择方法*

针对软件缺陷数据集中不相关特征和冗余特征会降低软件缺陷个数预测模型的性能的问题,提出了一种面向软件缺陷个数预测的混合式特征选择方法-HFSNFP。首先,利用ReliefF算法计算每个特征与缺陷个数之间的相关性,选出相关性最高的m个特征；然后,基于特征之间的关联性利用谱聚类对这m个特征进行聚类；最后,利用基于包裹式特征选择思想从每个簇中依次挑选最相关的特征形成最终的特征子集。实验结果表明,相比于已有的五种过滤式特征选择方法,HFSNFP方法在提高预测率的同时降低了误报率,且G-measure与RMSE度量值更佳；相比于已有的两种包裹式特征选择方法,HFSNFP方法在保证了缺陷个数预测性能的同时可以显著降低特征选择的时间。

A hybrid feature selection method for the number of software faults prediction

Focused on the issue that the irrelevant and redundant features in software defect data degrade the performance of the number of software faults prediction models, this paper proposed a Hybrid Feature Selection method for the Number of Faults Prediction (HFSNFP) . Firstly, HFSNFP computes the relevance between every feature and the number of fault with ReliefF algorithm and selects the top m most relevant features. Then, HFSNFP groups the m features with spectral clustering algorithm according to the correlation between every two features; Finally, HFSNFP selects the most relevant features from each resulted cluster to form the final feature subset using a wrapper search. Compared with five existing filter-based feature selection methods, the experimental results show that HFSNFP increases PD value, reduces PF value and achieves better G-measure and RMSE values. Comparison with two wrapper-based feature selection methods demonstrates that HFSNFP can achieve the high performance of the number of faults prediction and reduce the running time of feature selection.