Software defect prediction has been studied as an important research topic for 30 years in software en- gineering. Recently, with the development of machine learning techniques, traditional machine learning has been applied in software defect prediction based on static code attributes successfully. However, the traditional machine learning does not consider the cost-sensitive problem and class-imbalance problem in software defect prediction appli- cations. We study the application of cost-sensitive neural networks based on over-sampling and threshold-moving to software defect prediction. The experimental results on NASA software defect prediction benchmarking dataset dem- onstrate the algorithm's efficacy.%软件缺陷预测作为软件工程领域的重要研究内容已有近30年。近年来,随着机器学习技术的发展,传统机器学习技术基于静态代码属性的软件缺陷预测领域得到广泛应用。然而,传统的机器学习算法并未考虑软件缺陷预测过程中,常见的代价敏感问题与类不均衡问题。文中将基于过采样技术和阈值移动技术的代价敏感神经网络算法应用于软件缺陷预测领域,从而解决该领域的代价敏感问题与类不均衡问题。在NASA软件缺陷预测标准数据集上的实验证明了其有效性。
展开▼