Entre métricas e percepção: uso de aprendizado de máquina e LLMs na detecção e refatoração de code smells

Abstract

Software quality is essential for maintainability, readability, testability, and the long-term evolution of systems, with a large portion of maintenance costs related to understanding existing code. In this context, code smells emerge as indicators of structural or stylistic issues that degrade architecture and introduce technical debt. While metric-based detectors are widely used, they face limits of interpretability and generalization, whereas recent advances with machine learning (ML) algorithms and large language models (LLMs) expand automation possibilities but might introduce anomalies. This dissertation investigates how artificial intelligence techniques can support software quality improvement by connecting objective ML signals with human perception in refactorings produced by LLMs. The study was organized in three stages. In the first, five supervised classifiers were applied to four smells identified in systems from Qualitas Corpus. Tree-based models performed better and surpassing accuracy of 96% establishing replicable baselines. In the second stage, we propose TwinCode a open source tool developed to support empirical experiments with side by side code comparison and structured questionnaires. The tool was validated with 12 participants who reported high usability and consistency. In the third stage, we examined the developers’ perceptions of LLM-based refactorings through blind comparisons and interviews. In 97% of choices, intervieews chose the refactored versions and associated their choice with improvements in readability, modularity, and maintainability, even though they reported over-engineering issues. Our results demonstrate that: (i) ML algorithms provide robust signals for code smell detection, (ii) TwinCode contributes to the methodological standardization of empirical studies, and (iii) LLMs have potential to support refactorings when applied with critical supervision.Taken together, this dissertation integrates objective metrics and human judgments, offering evidence applicable to both academical and professional software development.