基于重子节点抽象语法树的软件缺陷预测

在实际软件项目开发过程中，软件缺陷预测能辅助测试人员找到项目中可能存在缺陷的位置，并通过抽象语法树（AST）获取项目模块中隐藏的结构和语义信息，此类信息有助于提高缺陷预测精度。提出基于重子节点抽象语法树的缺陷预测方法，在提取节点信息时保留节点的类型信息和对应代码语义的值信息，并使用特殊字符串代替没有值信息的节点。通过树链剖分思想将AST分割为重子节点和轻子节点，优先选择重子节点作为序列化向量中的节点，同时利用深度学习网络学习节点序列中的源代码结构和语言实现软件缺陷预测。实验结果表明，与DFS方法相比，该方法在基于注意力机制的循环神经网络深度学习模型上的F1值和AUC值平均提升约3%和4%，具有更好的缺陷预测效果。

Software Defect Prediction via Heavy Son Node-based Abstract Syntax Tree

In the actual software project development, software defect prediction can assist testers to find possible defects in the project.Through the Abstract Syntax Tree(AST), the hidden structure and semantic information in the project module can be obtained, which helps to improve the accuracy of defect prediction.This paper proposes a defect prediction method using Heavy Son(HS) node-based abstract syntax tree.In node information extraction, the type information of the node and the value information of the corresponding code semantics are retained, and the nodes without value information are replaced with a special string.Then by using the idea of tree chain division, the AST is divided into HS nodes and Light Son(LS) nodes.The HS nodes are selected in preference as the nodes in serialized vectors.At the same time, the deep learning network is used to learn the source code structure and language in the node sequence to realize software defect prediction.Experimental results show that compared with the DFS method, the proposed method improves the F1-measure by 3% and the AUC value by 4%, has a better defect prediction effect.