基于注意力和代价敏感的软件缺陷预测方法

Software Defect Prediction Method Based on Attention and Cost Sensitivity

摘要:

软件缺陷预测的目的是预先识别容易出现缺陷的代码模块以帮助软件质量保障团队适当的分配资源和人力；当前基于稳定学习的软件缺陷预测方法在特征提取过程中缺乏代码图像的全局信息，并忽视了不平衡数据对模型性能的影响；为了解决上述问题，论文提出了一种基于注意力和代价敏感的软件缺陷预测方法；该方法在SDP-SL的神经网络中增加了全局注意力模块，重点关注图像中和缺陷代码相关的特征，并将分类器的损失函数改进为代价敏感的损失函数，降低类不平衡对模型性能的影响；为了评估SDP-SLAC的性能，在PROMISE数据库中的10个开源Java项目上进行了多组比较实验；实验结果表明，SDP-SLAC方法可以有效提升缺陷预测模型的性能。

Abstract:

The purpose of software defect prediction is to identify code modules that may have defects in advance, assisting the software quality assurance team to allocate resources and manpower appropriately. Currently, the Software Defect Prediction method based on Stable Learning lacks global information of code images during the process of feature extraction and disregards the impact of imbalanced data on model performance. To address these issues, a Software Defect Prediction method based on Attention and Cost sensitivity is proposed here. This method enhances the neural network of SDP-SL with a global attention module that focuses on the features related to defective code in the images. Moreover, it improves the classifier's loss function to a cost sensitive loss function, mitigating the impact of class imbalance on model performance. To evaluate the performance of SDP-SLAC, multiple comparative experiments were conducted on ten open-source Java projects from the PROMISE database. The results unveil that the SDP-SLAC method effectively enhances the performance of defect prediction models.