Background: Code smells indicate underlying quality issues that negatively impact software maintainability. Refactoring is a common way to improve code quality by restructuring it, often removing these code smells. While many recommendations exist on how to refactor code smells, we do not fully understand how developers how they are removed by developers in the real world. Aim: In this study, we aim to investigate the evolution of code smells and the impact of applied refactoring techniques. Method: Our study addresses this gap by investigating both implementation and design smells and the refactoring techniques developers use to remove them. We also explore how often code smells are removed using established refactoring techniques. We analyzed 212,664 commits from 87 open-source Java projects using both automated tools and manual review to understand the relationship between code smells and refactoring. Results: Our key findings include: a) Extract method refactoring is most effective at fixing multiple smell types, b) Most applied refactorings do not remove code smells, c) About 82% of removed code smells are “dangling” i.e., they are removed without a matching refactoring technique, and d) Design smells typically last longer in codebases than implementation smells. Conclusions: This research improves our understanding of the interplay between code smells and refactoring effectiveness. Our results can help researchers develop better tools and guide software engineers in making their refactoring processes more efficient.

The use of machine learning to automate the detection of refactoring candidates is a rapidly evolving research area. The majority of work in this direction uses source code metrics and commit messages to predict refactoring candidates and do not exploit the rich semantics of source code. This paper proposes a new approach for extract method refactoring candidates identification. First, we propose a novel mechanism to identify negative samples for the refactoring candidate identification task. We then employ a self-supervised autoencoder to acquire a compact representation of source code generated by a pre-trained large language model. Subsequently, we train a binary classifier to predict extract method refactoring candidates. Experiments show that our new approach outperforms the state of the art by 30% in terms of F1 score. The proposed work has implications for researchers and practitioners. Software developers may use the proposed automated approach to predict refactoring candidates better. This study will facilitate the development of improved refactoring candidate identification methods that the researchers in the field could use and extend.