Improving the identification of code smells by combining structural and semantic information

Abstract

In software engineering, a code smell is an indication of a deeper problem in the source code, hindering the maintainability and evolvability of the system. In the literature, there is a significant emphasis on the detection of code smells because of its importance as a maintenance task. Most of previous studies focus in their analyses on one source of information, i.e. structural, historical or semantic information. However, some instances of bad smells could be identified by a type of information but missed by another one. In this paper, we propose an improved detection approach that combines structural and semantic information in order to fully exploit their complementarity in the identification of code smells. Both information are extracted separately using conventional and deep learning methods. For the evaluation, we have selected five open source projects which are JHotDraw, Apache Karaf, Freemind, Apache Nutch and JEdit. In order to optimize our performance results, we have set up four different experiments and compare between them. The obtained accuracy results confirm the effectiveness of combining structural and semantic information in improving the detection of code smells.