Can Lexicon Bad Smells Improve Fault Prediction?

In software development, early identification of fault-prone classes can save a considerable amount of resources. In the literature, source code structural metrics have been widely investigated as one of the factors that can be used to identify faulty classes. Structural metrics measure code complexity, one aspect of the source code quality. Complexity might affect program understanding and hence increase the likelihood of inserting errors in a class. Besides the structural metrics, we believe that the quality of the identifiers used in the code may also affect program understanding and thus increase the likelihood of error insertion. In this study, we measure the quality of identifiers using the number of Lexicon Bad Smells (LBS) they contain. We investigate whether using LBS in addition to structural metrics improves fault prediction. To conduct the investigation, we assess the prediction capability of a model while using i) only structural metrics, and ii) structural metrics and LBS. The results on three open source systems, ArgoUML, Rhino, and Eclipse, indicate that there is an improvement in the majority of the cases.