Automatically recommending components for issue reports using deep learning

Abstract

Today’s software development is typically driven by incremental changes made to software to implement a new functionality, fix a bug, or improve its performance and security. Each change request is often described as an issue. Recent studies suggest that a set of components (e.g., software modules) relevant to the resolution of an issue is one of the most important information provided with the issue that software engineers often rely on. However, assigning an issue to the correct component(s) is challenging, especially for large-scale projects which have up to hundreds of components. In this paper, we propose a predictive model which learns from historical issue reports and recommends the most relevant components for new issues. Our model uses Long Short-Term Memory, a deep learning technique, to automatically learn semantic features representing an issue report, and combines them with the traditional textual similarity features. An extensive evaluation on 142,025 issues from 11 large projects shows that our approach outperforms one common baseline, two state-of-the-art techniques, and six alternative techniques with an improvement of 16.70%–66.31% on average across all projects in predictive performance.