Software Fault Localization Based on Network Spectrum and Graph Neural Network

Abstract

Accurate fault localization renders software test resource allocation and maintenance cost-efficient. However, this is challenging when there are false alarm repercussions caused by module coupling of complex software. In this article, therefore, we propose a new method for multiple software fault localization from the perspective of network spectrum based on a graph neural network model. First, we constructed the network model of the software under test to represent the coupling relationships among software modules based on complex network theory. In addition, test suits were executed and recorded to construct the program spectrum. Subsequently, the software network and program spectrum were fused into the network spectrum, and we reprocessed it with feature dimension reduction, normalization, and graph-based class-imbalance treatment. The graph neural network was then used to construct a multiple-fault location model based on the processed network spectrum. Empirical studies were performed on the Defects4J dataset. The experimental results indicated that the proposed method outperformed six baseline methods (with an average improvement of 13.03% on the T-EXAMscore). This study is expected to provide insights into more smart software quality and reliability assurance.