Code Smell Detection using Hybrid Machine Learning Algorithms

Code smells frequently leads to the discovery of decreased code quality, drains on application resources, or even critical security vulnerabilities embedded within the application's code. While code smells may not always indicate a particularly serious problem, it do often lead to the discovery of these issues. Software's structural characteristics lead to a design issue that makes it challenging to manage and maintain code refactoring. The goal of the current research is to create methods for identifying code smells. The machine learning algorithm is a reliable method for individualized smell detection, but there aren't many studies on how well it works for different developers. In this proposed work used two different deep learning algorithms and five different machine learning ensembles to detect suspicious code. Investigation of the Data class, God class, Feature-envy, and Long-method datasets revealed that each contained various levels of code smells. Although there is room for improvement, the outcomes of prior publications' applications of machine learning and stacking ensemble learning methods to this dataset were satisfactory. A class balancing method (SMOTE) was implemented to address the problem of class imbalance within the datasets. While the Feature- envy dataset with the selected dozen metrics produced the lowest accuracy (91.45%) for the Max voting method, the Long-method dataset with the various chosen metrics produced the highest accuracy (100%) for all five methods.