Many learning based fault localization approaches haven been proposed to improve the effectiveness by fusing various dimension of fault diagnosis features. However, method calls behavior has been neglected, and the interaction between features has not been fully explored. To solve this problem, firstly, a fault localization method by mining software behavior graphs has been proposed to improve the effectiveness of localizing function call related faults. Then, a fault localization approach by wide & deep learning on multi-feature groups has been proposed. Not only the spectrum based and mutation based suspiciousness features have been analyzed, but also the behavior based and invariants based suspiciousness, the static metrics, as well as the combined features of crash stack trace with the invariants change features have been integrated. Wide & Deep model is adopted as the ranking model, to explore the relationships between these features, so as to improve the effectiveness of fault localization. Experiments on practical software defects benchmark Defects4J have shown that our model outperforms the traditional spectrum-based and mutation-based approaches, it also outperforms the state-art-of learning-based approaches on the capability of early fault detection.

已经提出了许多基于学习的故障定位方法，以通过融合各种维度的故障诊断特征来提高有效性。然而，方法调用行为被忽略了，特征之间的交互也没有得到充分的探索。针对这一问题，首先提出了一种挖掘软件行为图的故障定位方法，以提高定位函数调用相关故障的有效性。然后，提出了一种对多特征组进行广泛和深度学习的故障定位方法。不仅分析了基于频谱和基于变异的可疑性特征，而且还集成了基于行为和基于不变性的可疑性、静态度量以及崩溃堆栈跟踪与不变性变化特征的组合特征。采用Wide & Deep模型作为排序模型，探索这些特征之间的关系，从而提高故障定位的有效性。在实际软件缺陷基准 Defects4J 上的实验表明，我们的模型优于传统的基于频谱和基于突变的方法，它在早期故障检测能力方面也优于基于学习的最新方法。