基于迁移学习的软件缺陷预测方法研究

On Software Defect Prediction Method Based on Transfer Learning

针对传统的软件缺陷预测方法难以在单独的项目中利用小规模训练数据的问题,提出了一种基于迁移学习的软件缺陷预测技术,利用已有的项目辅助新项目的预测.该方法在源项目和目标项目之间寻找一个公共特征空间,使得在公共特征空间上2个项目的数据分布距离最小.在这个公共空间进行模型训练,以达到迁移分类的效果.实验结果显示该方法相对传统的缺陷预测算法有更好的预测性能,并且充分利用了原始训练数据,可以更高效地运用于各种软件缺陷预测任务.

To solve the problem of lacking training samples in a new proj ect,a new defect prediction meth-od based on transfer learning has been presented.Instead of training models only on the new proj ect,his-torical proj ect data sets with similarities are used to train the model.A sub-space is learned between source and proj ect first.Then the prediction model is trained on the sub-space which is used to predict defects for new proj ects.The experimental results show that the new method has better performance than normal de-fect prediction methods,which means the new method can be efficiently used in defect prediction tasks in the software engineering.