Research on significance metrics for software change-proneness prediction

Since existing software metrics do not consider the whole structure of a software project in evaluating classes, two new static metrics based on significance are proposed to build a better software change-proneness prediction model. The new metrics represent a project as a graph with classes being vertices and dependencies between classes being edges. The importance of a class is evaluated in the whole structure of the project, and the significance metrics for software change-proneness prediction are then proposed. Six open source software projects are used to validate the new metrics. Data used in experiments include two parts. The first part is metric of each class, including 4 volume metrics, 6 complexity metrics and 2 significance metrics, and the second part is the number of changed lines of each class during the selected history period. Experimental results show that when the new metrics are combined with other metrics, the accuracy increases on 5 projects by 1.16% on average, and the area under ROC curve increases on 6 projects by 3.65% on average.