An Imbalanced Deep Learning Model for Bug Localization

Abstract

Debugging and locating faulty source files are tedious and time-consuming tasks. To improve the productivity and to help developers focus on crucial files, automated bug localization models have been proposed for years. These models recommend buggy source files by ranking them according to their relevance to a given bug report. There are two significant challenges in this research field: (i) narrowing the lexical gap between bug reports which are typically described using natural languages and source files written in programming languages; (ii) reducing the impact of imbalanced data distribution in model training as a far fewer of source files relate to a given bug report while the majority of them are not relevant. In this paper, we propose a deep neural network model to investigate essential information hidden within bug reports and source files through capturing not only lexical relations but also semantic details as well as domain knowledge features such as historical bug fixings, code change history. To address the skewed class distribution, we apply a focal loss function combining with a bootstrapping method to rectify samples of the minority class within iterative training batches to our proposed model. We assessed the performance of our approach over six large scale Java open-source projects. The empirical results have showed that the proposed method outperformed other state-of-the-art models by improving the Mean Average Precision (MAP) and Mean Reciprocal Rank (MRR) scores from 3% to 11% and from 2% to 14%, respectively.