BugPre: an intelligent software version-to-version bug prediction system using graph convolutional neural networks

Abstract Since defects in software may cause product fault and financial loss, it is essential to conduct defect prediction (SDP) identify the potentially defective modules, especially early stage of development lifecycle. Recently, cross-version (CVDP) began draw increasing research interests, employing labeled data prior version within same project predict current version. As a dynamic process, distribution (such as defects) during change get changed. Recent studies utilize machine learning (ML) techniques detect defects. However, due close dependencies between updated unchanged code, ML-based methods fail model long deep dependencies, causing high false positive. Furthermore, traditional detection performed on entire project, efficiency relatively low, large-scale projects. To this end, we propose BugPre , CVDP approach address these two issues. novel framework that only conducts efficient changed modules utilizes variable propagation tree-based associated analysis method obtain Besides, constructs graph leveraging code context dependences uses convolutional neural network learn representative characteristics thereby improving capability when changes occur. Through extensive experiments open-source Apache projects, experimental results indicate our outperforms three state-of-the-art approaches, F1-score has increased by higher than 16%.