BugRadar: Bug localization by knowledge graph link prediction

Abstract

: Information Retrieval-based Bug Localization (IRBL) aims to design automatic systems that find buggy files according to bug reports, which can reduce the time consumption to fix bugs for programmers. There has been extensive research on IRBL techniques in recent years. However, these methods cannot make full use of the structure information in bug reports and source files. : In this paper, we propose a novel scheme BugRadar. It combines text features and structure features from bug reports and source files for bug localization. Especially, BugRadar leverages a knowledge graph to make use of structure features. : We originally propose a knowledge graph named TriGraph based on structure features and apply hyperbolic attention embedding to get the link prediction scores. For text features, we propose Partial Text Similarity which improves traditional Text Similarity and Method Level Text Similarity. We also propose Word Collaborative Filtering Score which leverages historical bug reports with more attention on important terms. Finally, we calculate the final suspicious scores based on the structure features, text features, and fixing time information from bug fixing history with a neural network. : We apply our scheme to four projects (Tomcat, SWT, JDT, and Birt) in a popular dataset and get approving results. BugRadar gets better results than other state-of-the-art methods on three projects out of the four. It achieves a relative improvement of 8.8% in SWT and 9.8% in JDT for Mean Average Precision compared to the previous best scheme KGBugLocator and 11.4% in Birt compared to Adaptive Regression. : BugRadar can achieve approving performance on large-scale projects with enough historical bug reports. It verifies that knowledge graphs are capable of representing the structure features for bug localization. The novel Partial Text Similarity and Word Collaborative Filtering Score are both effective improvements for using text features.