An Empirical Framework for Code Smell Prediction using Extreme Learning Machine

Abstract

The software containing code smells indicates the violation of standard design and coding practices by developer during the development of the software system. Recent empirical studies observed that classes having code smells have higher probability of change proneness or fault proneness with respect to classes having no code smells [1]. The effort of removing bugs due to code smells increases exponentially if the smells are not identified during the earlier phases of software development. The code smell prediction using source code metrics can be used in starting phases of software development life cycle to reduce the maintenance and testing effort of software and also help in improving the quality of the software. The work in this paper empirically investigates and evaluates different classification techniques, feature selection techniques, and data sampling techniques to handle imbalance data in predicting 7 different types of code smell. The conclusion of this research is assessed over 629 application packages. The experimental finding confirms the estimating capability of different classifiers, feature selection, and data imbalance techniques for developing code smell prediction models. Our analysis also reveals that the models developed using one technique are superior than the models developed using other techniques.