A Comparative Study of Different Strategies for Predicting Software Quality

Various methods have been developed for improving the quality of a software product, especially for high-assurance and missioncritical software systems. One commonly used approach is software quality modeling, in which software practitioners utilize software metrics and defect data collected during the software development process to build defect prediction models that will help to find poor-quality program modules. Those modules predicted to be fault-prone will receive more inspection and testing, thereby improving their quality. Efficacy of defect prediction models is influenced by relevance between software metrics and fault data. Usually not all software metrics in data repositories contribute equally to the occurrence of faults. Choosing the most important metrics (features) prior to the model training process is needed to improve the effectiveness of defect predictors. In this paper, we study 18 filter-based feature selection techniques and evaluate their effectiveness through a case study performed on 16 different software data sets. Among the 18 techniques, six of them are standard filter-based methods, while 11 of them are threshold-based feature selection (TBFS) techniques proposed by our research team recently. The last one is based on signal to noise ratio (S2N), which is a widely used concept in electrical and communication engineering, but which is rarely used in feature selection. The experimental results demonstrate that the TBFS techniques perform similarly to the standard techniques and the S2N technique shows significantly better performance than the other 17 approaches.