| Summary: | Context: Lines of code (LOC) is a fundamental software code measure that is widely used as a proxy for software development effort or as a normalization factor in many other software-related measures (e.g., defect density). Unfortunately, the problem is that it is not clear which lines of code should be counted: all of them or some specific ones depending on the project context and task in mind? Objective: To design a generator of task-specific LOC measures and their counters mined directly from data that optimize the correlation between the LOC measures and variables they proxy for (e.g., code-review duration). Method: We use Design Science Research as our research methodology to build and validate a generator of task-specific LOC measures and their counters. The generated LOC counters have a form of binary decision trees inferred from historical data using Genetic Programming. The proposed tool was validated based on three tasks, i.e., mining LOC measures to proxy for code readability, number of assertions in unit tests, and code-review duration. Results: Task-specific LOC measures showed a “strong” to “very strong” negative correlation with code-readability score (Kendall’s <inline-formula> <tex-math notation="LaTeX">$\tau $ </tex-math></inline-formula> ranging from −0.83 to −0.76) compared to “weak” to “strong” negative correlation for the best among the standard LOC measures (<inline-formula> <tex-math notation="LaTeX">$\tau $ </tex-math></inline-formula> ranging from −0.36 to −0.13). For the problem of proxying for the number of assertions in unit tests, correlation coefficients were also higher for task-specific LOC measures by ca. 11% to 21% (<inline-formula> <tex-math notation="LaTeX">$\tau $ </tex-math></inline-formula> ranged from 0.31 to 0.34). Finally, task-specific LOC measures showed a stronger correlation with code-review duration than the best among the standard LOC measures (<inline-formula> <tex-math notation="LaTeX">$\tau $ </tex-math></inline-formula> = 0.31, 0.36, and 0.37 compared to 0.11, 0.08, 0.16, respectively). Conclusions: Our study shows that it is possible to mine task-specific LOC counters from historical datasets using Genetic Programming. Task-specific LOC measures obtained that way show stronger correlations with the variables they proxy for than the standard LOC measures.
|
|---|