Do code data sharing dependencies support an early prediction of software actual change impact set?

摘要

Existing studies have shown that structural dependencies within code are good predictors for code actual change impact set—a set of entities that repeatedly changing together to ensure a consistent and complete change. However, the result is far from ideal, particularly when insufficient historical data are available at an early stage of software development. This paper demonstrates that a better understanding of data dependencies in addition to call dependencies greatly improves actual change impact set prediction. We propose a new approach and tool (namely, CHIP) to predict software actual change impact sets leveraging both call and data sharing dependencies. For this purpose, CHIP employs novel extensions (dependency frequency filtering and shared data type idf filtering) to reduce false positives. CHIP assumes that developers know initial places where to start making changes in the source code even though they may not know all changes. This approach has been empirically evaluated on 4 large‐scale open source systems. Our evaluation demonstrates that data sharing dependencies have a complementary impact on software actual change impact set prediction as compared with predictions based on call dependencies only. CHIP improves the F2‐score compared with the predictors using both Program Dependence Graph and evolutionary couplings.