Two-tier deep and machine learning approach optimized by adaptive multi-population firefly algorithm for software defects prediction

Software plays a progressively crucial role, where automated software systems control essential operations. Since development needs also progressively expand, manual code reviews become increasingly difficult, frequently resulting in testing lasting longer than development itself. An encouraging option for enhancing defect identification within the source code involves combining artificial intelligence and natural language processing (NLP). Analyzing source code offers an efficient approach to enhance defect detection and prevent errors in the code. This study investigates source code analysis using NLP and machine learning, where traditional and contemporary techniques of error detection are evaluated. Metaheuristics algorithms are utilized to tune machine learning classifiers, and an altered variant of the well-known firefly algorithm is proposed as part of this research. A two-tier framework is suggested, consisting of a convolutional neural network (CNN), which handles complex feature spaces, while eXtreme gradient boosting (XGBoost), adaptive boosting (AdaBoost), and categorical boosting (CatBoost) classifiers are employed within the second-tier for improving defect detection. Supplementary simulations employing custom term frequency inverse document frequency encoding are also executed to showcase the capabilities of the suggested framework. In total, seven experiments are carried out with publicly accessible datasets. The accuracy of CNN is 80.6% for the defect prediction task, which is enhanced with the second layer using XGBoost, AdaBoost, and CatBoost to nearly 81.5%. The experiments with the NLP approach exhibit superior outcomes, where XGBoost, AdaBoost, and CatBoost achieve accuracies of 99.6%, 99.7%, and 99.8%, indicating the large potential of the suggested approach in the software testing domain.