基于动态分布适应网络的跨项目缺陷预测

Cross-Project Defect Prediction based on Dynamic Distributed Adaptive Networks

摘要:

在软件缺陷预测中，跨项目缺陷预测是基于源项目的标记数据来训练模型，并预测当前正在开发的目标项目的缺陷；然而，两个不同项目数据之间的分布差异往往限制了跨项目缺陷预测模型的能力；由于源域和目标域的数据通常来自不同的分布，因此现有方法主要集中于适应跨域边缘或条件分布；在实际应用中，现有方法无法定量评估边缘分布和条件分布的重要性，这将导致传输性能不理想；论文提出了一种基于动态分布适应网络的跨项目缺陷预测方法来解决分布差异问题，它利用迁移学习能够定量评估每个分布的相对重要性；论文对来自3个公共数据集的 24个项目进行了实验，以验证所提出的方法。结果表明，平均而言在 AUC 和 F1 分数上分别比所有基线方法高出至少 1.3% 和 5.7%。这表明所提出的方法具有良好的性能特点。

Abstract:

In software defect prediction, cross-project defect prediction is based on the labeled data from a source project to train a model and predict defects in the target project currently under development. However, the distribution differences between data from two different projects often limit the ability of cross-project defect prediction models. As the data from the source domain and target domain typically come from different distributions, existing methods mainly focus on adapting cross-domain marginal or conditional distributions. In practical application, existing methods are unable to quantitatively evaluate the importance of marginal and conditional distributions, which leads to suboptimal transfer performance. This paper proposes a cross-project defect prediction method based on a dynamic distribution adaptation network to address the distribution difference issue, utilizing transfer learning to quantitatively evaluate the relative importance of each distribution. The paper conducts experiments on 24 projects from 3 public datasets to validate the proposed method. The results show that, on average, the proposed method outperforms all baseline methods by at least 1.3% and 5.7% in terms of AUC and F1 scores, respectively. This indicates that the proposed method exhibits good performance characteristics.