Efficient feature envy detection and refactoring based on graph neural network

As one type of frequently occurring code smells, feature envy negatively affects class cohesion, increases coupling between classes, and thus hampers software maintainability. While progress has been made in feature envy detection, two challenges still persist. Firstly, existing approaches often underutilize method call relationships, resulting in suboptimal detection efficiency. Secondly, they lack the emphasis on feature envy refactoring, which is however the ultimate goal of feature envy detection. To address these challenges, we propose two approaches: SCG (SMOTE Call Graph) and SFFL (Symmetric Feature Fusion Learning). SCG transforms the feature envy detection problem into a binary classification task on a method call graph. It predicts the weights of edges, termed calling strength, to capture the strength of method invocations. Additionally, it converts the method-method call graph into a method-class call graph and recommends the smelly method to the external class with the highest calling strength. As a holistic approach focusing on refactoring feature envy directly, SFFL leverages four heterogeneous graphs to represent method-class relationships. Through Symmetric Feature Fusion Learning, it obtains representations for methods and classes. Link prediction is then employed to generate the refactored method-class ownership graph, which is regarded as the refactored results. Moreover, to address the limitations of existing metrics in accurately evaluating refactoring performance, we introduce three new metrics: precision2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textit{precision}_2$$\end{document}, recall2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textit{recall}_2$$\end{document} and F1-score2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textit{F}_1\text {-score}_2$$\end{document}. Extensive experiments on five open-source projects demonstrate the superiority of SCG and SFFL. The code and dataset used in our study are available at https://github.com/HduDBSI/SCG-SFFL.