COSTAR: Software Code Smell Detection Through Tree-Based Abstract Representation

doi:10.1109/TR.2025.3648404

Return

COSTAR: Software Code Smell Detection Through Tree-Based Abstract Representation

2026-01-01

PRE

Mahipal Jadeja

Satyendra Singh Chouhan

Santosh Singh Rathore

DOI：10.1109/TR.2025.3648404

Original

Original request for help

Save

Abstract

En 中文

Code smells are suboptimal code structures that increase software maintenance costs and are challenging to detect manually. Researchers have explored automatic code smell detection using machine learning (ML) methods, which rely heavily on static code metrics or source code representation. Static code metrics often rely on structural attributes, such as lines of code, cyclomatic complexity, or comment density. However, these metrics do not always reflect true code complexity and provide only quantitative insights without inherently detecting poor coding practices. In contrast, representations such as abstract syntax trees (ASTs) focus on the structural and syntactic elements of code, capturing hierarchical and contextual relationships within the source code. This enables precise identification of code structures such as loops, function calls, and conditionals, which are essential for detecting code smells. This article introduces code smell detection through tree-based abstract representation (COSTAR), a source code representation technique using ASTs to uniquely represent each source code instance. COSTAR captures the hierarchical structure of the source code by extracting all paths from the root to individual nodes within the AST. By employing a pretrained sentence bidirectional encoder representations from transformers embedding model, COSTAR generates vectors for each extracted path. The subsequent calculation of the mean of these vectors yields a precise and comprehensive source code representation. Extensive experiments were conducted to validate COSTAR's performance using various ML techniques on four benchmark MLCQ code smell datasets: Data Class, God Class (Blob), Feature Envy, and Long Method. Various performance metrics have been employed to evaluate the model's performance. The experimental results indicate that COSTAR enhances the performance of the code smell detection model compared to existing methods. An improvement in the F1-score ranging from 0.03 (Long Method) to 0.19 (Feature Envy) was observed. Furthermore, a comparison of COSTAR with state-of-the-art methods demonstrated that it outperformed approaches such as Code2Vec and CuBERT in code smell detection.

Keywords:

Code smell prediction

empirical study

imbalance learning

software maintainability

source code representation