Application of Deep Learning models for Code Smell Prediction

Software developer always strives for quality of the software as it tends to be more robust and easier to maintain. Code smells play as a hinder to the quality of the software as they are the surface indication of deeper problem in the source code and hence required to be removed as early as possible. Refactoring is one of the methods to improve the quality of the software without affecting its external behavior. It removes bad smells present in the code as they are surface indication of deeper problem which may lead towards the failure of the software. In this work, we have proposed an approach to predict code smells in the source code using Deep learning algorithms. More specifically, we have trained three deep learning algorithms (CNN, LSTM and MLP) with open-source dataset. Information Gain feature selection algorithm is also applied to get the most prominent attributes for the code smell prediction. After evaluating the performance of these algorithms, results shows that CNN outperforms all other deep learning algorithms with the accuracy of 99% on the selected dataset. This study would be useful for software developer team as they can utilize CNN algorithm in code smell prediction. Results would also help in predicting the code smell which in turn will help in the removal of these smells to enhance the quality of the software.