基于图可解释网络的软件错误定位

摘要:

软件错误定位技术旨在通过挖掘程序与测试用例执行数据，提升定位准确性。针对SBFL技术过于依赖二进制覆盖信息的问题，提出一种基于图可解释网络的软件错误定位方法，将测试执行转化为图结构，利用图注意网络建模深度挖掘代码片段隐含的信息及其相互关系，并采用强化学习思想对图注意力网络学习后的决策过程进行解释，从而确定关键节点，缩小错误定位范围。实验的场景设立在Defects4j数据集的5个项目进行了，并与SBFL及未经过解释的深度学习方法进行了对比。结果显示，基于图可解释网络的定位方法在Top-1、Top-3和Top-5指标上分别提升了7.26%、7.56%和9.96%，EXAM指数也提升了8.98%，显著优于其他方法。

Abstract:

Software fault localization aim to improve localization accuracy by mining program and test cases execution data. Addressing the issue that SBFL technology relies too much on binary coverage information, a software fault localization method based on graph-interpretable networks is proposed. This method transforms test execution into a graph structure and utilizes graph attention networks to model and deeply mine the implicit information and interrelationships among code segments. It also applies reinforcement learning principles to explain the decision-making process after graph attention network learning, thereby determining key nodes and narrowing the fault localization range. The experiments were conducted on five projects from the Defects4j dataset and compared with SBFL and uninterpreted deep learning methods. The results showed that the localization method based on graph-interpretable networks improved the Top-1, Top-3, and Top-5 metrics by 7.26%, 7.56%, and 9.96%, respectively, and the EXAM index also increased by 8.98%, significantly outperforming other methods.