On Transfer Learning in Code Smells Detection

EasyChair Preprint 7717

Abstract

The incidence of code smells is often associated with software quality degradation. Several studies present the importance of detecting and tackling the incidence of smells in the source code. However, existing technologies to detect code smells are dependent on the programming language. Consequently, several programming languages are largely employed by the software community without proper technologies code smell detection. This paper investigates the use of transfer learning to detect code smells in different programming languages. We selected five programming languages among the ten most used languages according to \textit{StackOverflow}: Java, C\#, C++, Python, and JavaScript. We selected open-source projects to obtain the datasets for training and testing. Results indicate high levels of effectiveness in detecting Complex Methods from other programming languages through transfer learning models, except for Python. This finding can help developers and researchers to apply the same code smell detection strategies in different programming languages. The results also indicate that the particular behavior observed with Python is partially due to key structural differences in this programming language.

Keyphrases: Transfer Learning, code smells, deep learning

@booklet{EasyChair:7717,
  author    = {Moabson Ramos and Rafael de Mello and Baldoino Fonseca},
  title     = {On Transfer Learning in Code Smells Detection},
  howpublished = {EasyChair Preprint 7717},
  year      = {EasyChair, 2022}}