基于代价敏感学习的软件缺陷预测方法

摘要

Software defect prediction is an important way to improve the quality of software development and raise the testing efficiency. In this paper,analyze the characteristics of software defect prediction and focus on the research of redundancy features and the imbalance class problem existed in current software defect. In order to solve the difficulty of software defect prediction caused by redundancy fea-tures in software modules,improving the accuracy for software defect prediction,adopt a new method named Cost-Censitive Laplacian Eigenmaps (CSLE) to reduce the dimensionality of original sample space, improving the distance measurement method of Laplacian Eigenmaps (LE) to enhance the dimension reduction mapping accuracy. In addition,propose a new method named Cost Sensitive Back Propagation Neural Network (CSBPNN) to classify the software module,adjusting the weights and bias parameters of BP neural net-work,which makes the error of BP neural network to flawed software modules points more sensitive,further improving the classification effect. Compared with the latest several software defect prediction methods on NASA software datasets,prove that this method can im-prove the recall rate and F-measure value in software defect prediction.%软件缺陷预测是改善软件开发质量、提高测试效率的重要途径.文中分析了软件缺陷预测的特点,同时针对当前软件缺陷预测中存在特征冗余问题和类不平衡问题进行了深入研究.首先为了解决软件模块中的特征冗余问题给软件缺陷预测造成困难,提高对软件缺陷预测的准确率,采用基于代价敏感的拉普拉斯特征映射方法(CSLE)对原样本空间进行降维,改进拉普拉斯算法(LE)中的距离度量方式,提高降维映射精度;然后通过基于代价敏感的神经网络的方法(CSB-PNN)对软件模块进行分类,调整BP神经网络的权值和偏置参数,使BP神经网络对有缺陷软件模块的误分更加敏感,进一步提高分类效果.在NASA软件缺陷标准数据集上与最新的几种软件缺陷预测方法相比,文中提出的方法能够有效提高有缺陷样本的召回率和F-measure值.