Authors:
Mouna Hadj-Kacem
and
Nadia Bouassida
Affiliation:
Mir@cl Laboratory, Tunisia
Keyword(s):
Hybrid Approach, Deep Learning, Auto-encoder, Artificial Neural Networks, Code Smell Detection.
Abstract:
The detection of code smells is a fundamental prerequisite for guiding the subsequent steps in the refactoring process. The more the detection results are accurate, the more the performance of the refactoring on the software is improved. Given its influential role in the software maintenance, this challenging research topic has so far attracted an increasing interest. However, the lack of consensus about the definition of code smells in the literature has led to a considerable diversity of the existing results. To reduce the confusion associated with this lack of consensus, there is a real need to achieve a deep and consistent representation of the code smells. Recently, the advance of deep learning has demonstrated an undeniable contribution in many research fields including the pattern recognition issues. In this paper, we propose a hybrid detection approach based on deep Auto-encoder and Artificial Neural Network algorithms. Four code smells (God Class, Data Class, Feature Envy an
d Long Method) are the focus of our experiment on four adopted datasets that are extracted from 74 open source systems. The values of recall and precision measurements have demonstrated high accuracy results.
(More)