Novel Insights on Cross Project Fault Prediction applied to Automotive Software

Defect prediction is a powerful tool that greatly helps focusing quality assurance efforts during development. In case of the availability of fault data from in a particular context, there are multiple different ways of using such fault predictions in practice. Companies like Google, Bell Labs, or Cisco make use of fault prediction, whereas its use within automotive industry has not yet gained a lot of attraction, although, modern cars require a huge amount of software to operate. In this paper, we want to contribute to adopt fault prediction techniques for automotive software projects. Hereby we rely on a publicly available data set comprising fault data from three automotive software projects. When learning a fault prediction model from the data of one particular project, we achieve a remarkably high and nearly perfect prediction performance for the same project. However, when applying cross-project prediction we obtain rather poor results. These results are rather surprising, because of the fact that the underlying projects are as similar as two distinct projects can possibly be within a certain application context. Therefore we investigate on the reasons behind this observation through correlation and factor analyses techniques. We further report the obtained findings and discuss the consequences for future applications of cross-project fault prediction in the domain of automotive software.