Severity Classification of Code Smells Using Machine-Learning Methods

Abstract

Code smell detection can be very useful for minimizing maintenance costs and improving software quality. Code smells help developers/programmers, researchers to subjectively interpret design defects in different ways. Code smells instances can have varied size, intensity or severity which needs to be focused upon as they affect the software quality accordingly. Therefore, this study aims to detect the severity of code smells from code smell datasets. The severity of code smells is significant for reporting code smell detection performance, as it permits refactoring efforts to be prioritized. Code smell severity also describes extent of effort required during software maintenance. In our work, we have considered four code smells severity datasets to detect the severity of code smell. These datasets are data class, god class, feature envy and long method code smells. This paper uses four machine-learning and three ensemble learning approaches to identify the severity of code smells. To improve the models’ performance, we used fivefold cross-validation method: Chi-square-based feature selection algorithm and parameter optimization techniques. We applied two-parameter optimization techniques, namely grid search and random search and also compared their accuracy. The conclusion of this study is that the XG Boost model obtained an accuracy of 99.12%, using the Chi-square-based feature selection technique for the long method code smell dataset. In this study, the results show that ensemble learning is best as compared to machine learning for severity detection of code smells. © 2023, The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd.