Uma aplicação do aprendizado por transferência na detecção de Code Smells

Abstract

During the software development, the presence of code smells has been related to the degradation of the software quality. Several studies present the relevance to detect smells in the source code and to apply refactoring. However, the existing approaches to detect code smells are limited to specific programming languages. In this context, this work aims to extend the techniques of code smell detection using learning transfer to build a large dataset for training and validation of machine learning models, using cataloged rules and extracted thresholds of the Designite tool. Collecting, then, a total of total 22,687, 8,501 e 5,953 detected smells in software projects of the respective programming languages, C++, Java e C#. In a sequence, we obtained 72 pretrained models and performed the transfer learning to evaluate the pre-trained model for smells in the dataset between programming languages. Our results revealed that the model RandomForest is the most appropriate to detect code smells like the Unecessary Abstraction design smell for the C# programming language. These results can help developers and researchers to apply the same code smell detection strategies for different programming languages and to apply the most appropriate training models for each code smell type and programming language.