基于不相似性的软件缺陷预测算法

Dissimilarity-Based Software Defect Prediction Algorithm

摘要:

软件缺陷预测是典型的类不均衡学习问题,其中有缺陷的样本数量远少于无缺陷的样本数量,但有缺陷的样本通常是预测的重点。现有的软件预测模型大多建立在基于静态度量元的软件缺陷数据集上,重点关注如何平衡类分布,而忽略了数据集中属性特征对软件缺陷的判别能力。当软件缺陷数据集中的属性特征对类目标概念缺乏判别能力时,传统机器学习算法难以构建有效的软件缺陷预测模型,从而无法获得有效的预测性能。为此,提出了一种基于不相似性的软件缺陷预测算法,通过改善软件缺陷数据集中属性的判别能力,进而提升软件缺陷预测性能。实验证明：基于不相似性的软件缺陷预测算法能够有效地改善传统机器学习算法在软件缺陷数据集上的预测性能。

Abstract:

Software defect prediction is a typical imbalance learning problem, in which the number of ‘True’ (defective) examples is far less than that of ‘False’ (Non-defective) ones. The minority class is often the learning objective. However, the existing researchers devote on building the prediction model on the static metric-based software defect data set, focusing on how to balance the imbalance class distribution, but ignoring the effect from the discrimination ability of features in the context of the software defect data set. Therefore, a dissimilarity-based software defect prediction algorithm (DSDPA) is proposed,which can be used for effectively increasing the performance of software defect prediction.