Assessing software defection prediction performance: why using the Matthews correlation coefficient matters

Abstract

A considerable body of empirical soware engineering research is dedicated to predicting which are the defect-prone components in soware systems [3, 9, 12, 16, 24]. e motivation is quite straightforward. If we know which parts of the system are likely to be problematic, we can allocate testing resources accordingly. Moreover, more recent studies also take into account varying component test costs e.g., depending on size [18]. Another important line of research is into the ability to use data from one soware project to make predictions for a dierent project [38].