Int J Performability Eng ›› 2020, Vol. 16 ›› Issue (6): 979-990.doi: 10.23940/ijpe.20.06.p16.979990

Previous Articles    

An HVSM-based GRU Approach to Predict Cross-Version Software Defects

Xue Baia, Hua Zhoub,*, and Hongji Yangc   

  1. a School of Software, Yunnan University, Kunming, 650000, China;
    b College of Big Data and Intelligent Engineering, Southwest Forestry University, Kunming, 650000, China;
    c School of Informatics, Leicester University, Leicester, LE1 7RH, England
  • Submitted on ; Revised on ; Accepted on
  • Contact: * E-mail address:
  • About author:Xue Bai is a doctoral student of the School of Software at Yunnan University, China. Her research interests include System Analysis and Integration and Software Engineering.
    Hua Zhou is a professor of the College of Big Data and Intelligent Engineering at Southwest Forestry University, China. His research interests include Information System Integration and Analysis.
    Hongji Yang is a professor at School of Informatics, Leicester University, England, His research interests include Creative Computing, Software Engineering and Internet Computing.

Abstract: Cross-version Software Defect Prediction (CSDP) can be used to predict defect-prone modules or the number of defects in the latest version by using defect rules learned from historical versions. This technique can not only reduce the cost of iterative development of software and improve the reliability of new version, but also be used to understand the causes of defects and improve the software development process. Recently, much effort has been paid to build accurate cross-version defect prediction models, including quality defect predictors, and develop modeling techniques. Liu et al. proposed a new way to construct the predictor based on software code metrics and software process metrics. This new predictor is called Historical Version Sequence of Metrics (HVSM), which is processed with RNN. However, their model has some shortcomings, such as the missing of evolution information in HVSM, difficulty in training RNN, and lack of the ability to predict the number of defects, etc. To solve these problems, we add software evolution metrics to HVSM and bring in a new deep learning technique, Gate Recurrent Unit (GRU), to enhance the HVSM. Our HVSM-based GRU model can predict both defect-prone modules and the number of defects. The experimental results show that the HVSM with incremental evolution metrics provide better performance and in most cases the HVSM-based GRU model outperforms the commonly used baseline models in CSDP.

Key words: cross-version defect prediction, HVSM, GRU, defect-prone prediction, defect number prediction