Evaluating Software Product Metrics with Synthetic Defect Data

Source code metrics have been used in past research to predict software quality and focus tasks such as code inspection. A large number of metrics have been proposed and implemented in consumer metric software, however, a smaller, more manageable subset of these metrics may be just as suitable for accomplishing specific tasks as the whole. In this research, we introduce a mathematical model for software defect counts conditioned on product metrics, along with a method for generating synthetic defect data that chooses parameters for this model to match statistics observed in empirical bug datasets. We then show how these synthetic datasets, when combined with measurements from actual software systems, can be used to demonstrate how sets of metrics perform in various scenarios. Our preliminary results suggest that a small number of source code metrics conveys similar information as a larger set, while providing evidence for the independence of traditional software metric classifications such as size and coupling.