A Systematic Review on Application of Deep Learning Techniques for Software Quality Predictive Modeling

Software quality prediction is the process of evaluating the software developed for various metrics like defect prediction, bug localisation, effort estimation etc. To evaluate these metrics a myriad of techniques have been developed in the literature, from manual assessment to application of machine learning and statistical testing. These methodologies, however, had lower accuracy in determining SQPMs due to their inability to model the complex relationships in the training data. With the wide emergence of deep learning, not only has the accuracy of the pre-existing models enhanced, but it has also opened doors for new metrics that could be evaluated and automated. This study performs a systematic literature review of research papers published from January 1990 to January 2019 that used deep learning to evaluate software quality prediction metrics (SQPM). The paper identifies 20 primary studies and 7 categories of application of deep learning in SQPM. Models using deep learning techniques significantly outperform other traditional methodologies in almost all studies. The concept and external threats to the models are limited, however the time taken to train these models is large. The techniques, currently predominantly applied for defect prediction, have shown promising results in other diverse software engineering fields like code search and effort estimation by modeling the source code efficiently. There is, hence, scope for incorporating deep learning further with pragmatic use and diverse application. The need to find scalable solutions, however, still persists.