Uma abordagem de deep-learning para realizar a predição de refatorações

Abstract

Research has shown that refactoring often leads to greater maintainability, resulting in more readable and understandable code for developers. However, when applying refactoring methods to increase software quality, Developers face challenges in identifying effective refactoring methods. It turns out that finding refactoring opportunities is a challenging task. One notable problem is the absence of specific, practical guidelines for determining the appropriate refactoring method for a given piece of code. Consequently, since decisions about when to refactor are often based on subjective concepts such as codesmells, Less experienced developers often rely on guidance from senior developers to determine when software needs to undergo refactoring. Previous research has shown that machine learning algorithms can be used to help developers identify refactoring opportunities. With recent advances in hardware, Deep learning algorithms have attracted more and more attention. In this research, we intend to evaluate the effectiveness of some Deep Learning models (CNN, RNN, LSTM and DenseLayer) in predicting refactoring opportunities, compared to traditional machine learning models. Specifically, We evaluate these models using standard metrics such as precision, recall, and accuracy. Our findings seem to suggest that although machine learning models generally outperform deep learning models, the latter perform better than the former when trained on unbalanced datasets.