基于局部稀疏重构度量学习的软件缺陷预测

Software Defect Prediction of Metric Learning Based on Local Sparse Reconstruction

随着计算机技术的不断发展,如何准确地预测出软件中潜在的缺陷显得至关重要。近年来,研究者们尝试把一些机器学习方法应用到软件缺陷预测领域中,但是这些方法在分类过程中大多使用了传统的欧氏距离。距离度量学习方法通过挖掘训练样本集的特征信息和标记信息,学习得到有效的距离度量,让样本在基于度量矩阵的新特征空间中具有更好的鉴别可分性。将距离度量学习方法引入到软件缺陷预测中,同时融入了局部稀疏重构信息,提出一种新的软件缺陷预测方法,即局部稀疏重构度量学习方法(LSRML)。该方法学习得到的距离度量具有很好的鉴别性,并有效地解决了噪声敏感问题。在软件工程NASA数据库上的实验结果表明,提出的方法具有较好的缺陷预测效果。

With the development of computer technology,how to predict the potential defects in software project preciously is an important topic. Recently, researchers have introduced some machine learning methods into the software defect prediction field. However, they usual- ly utilize the traditional Euclidean metric in classification phase. Distance metric learning can learn an effective distance metric by exploiting the feature and label information of training sets, which makes the original samples hold better discriminability in the new feature space. The distance mettle learning is introduced into the software defect prediction field, and a novel software defect prediction approach called Local Sparse Reconstruction based Metric Learning （LSRML） is proposed. It incorporates the local sparse reconstruction information into the distance metric learning scheme. The learned distance metric not only has favorable discriminability, but also effectively handles the noise problem. The experiment results on the NASA projects demonstrate the effectiveness of the proposed approach.