An effective change recommendation approach for supplementary bug fixes

Xin Xia, David Lo

Research output: Contribution to journalArticleResearchpeer-review

14 Citations (Scopus)


Bug fixing is one of the most important activities during software development and maintenance. A substantial number of bugs are often fixed more than once due to incomplete initial fixes which need to be followed up by supplementary fixes. Automatically recommending relevant change locations for supplementary bug fixes can help developers to improve their productivity. It also help improve the reliability of systems by highlighting locations that a developer potentially needs to change to completely remove a bug. Unfortunately, a recent study by Park et al. shows that many change recommendation techniques do not work for supplementary bug fixes. In this paper, to advance the capabilities of existing change recommendation techniques, we propose an effective approach named SupLocator to recommend relevant locations (i.e., methods) that need to be changed for supplementary bug fixes. Based on various relationships among methods, classes, and packages in the source code (such as containment, inheritance, historical co-change, etc.), SupLocator extracts six change relationship graphs. Next, SupLocator performs random walk on each of the 6 graphs, and for each it outputs a ranked list of candidate change locations. Finally, SupLocator combines these six ranked lists by leveraging genetic algorithm. To investigate the benefits of SupLocator, we perform experiments on three projects, i.e., Eclipse JDT, Eclipse SWT, and Equinox p2. The experimental results show that on average SupLocator can achieve top-1, top-5, and top-10 accuracies, mean reciprocal rank (MRR), and mean average precision (MAP) of 0.51, 0.65, 0.67, 0.58 and 0.32 for the three projects, which improve the best variants of the approach proposed by Park et al. by 1523.09, 639.70, 550.62, 919.41, and 1478.44 %, respectively. It also improves the approach proposed by Saul et al. in terms of top-1, top-5, and top-10 accuracies, MRR, and MAP by 71.81, 29.54, 18.30, 47.24, and 56.60 %, respectively. Statistical tests show that the improvements are statistically significant.

Original languageEnglish
Pages (from-to)455-498
Number of pages44
JournalAutomated Software Engineering
Issue number2
Publication statusPublished - 1 Jun 2017
Externally publishedYes


  • Change recommendation
  • Genetic algorithm
  • Random walk
  • Supplementary bug fixes

Cite this