Code Smell Detection Using Machine Learning in Static Analysis

Abstract

Code smells—recurring design or implementation symptoms that indicate deeper problems—degrade maintainability, increase fault-proneness, and inflate long-term costs. Traditional smell detection relies on heuristics woven into static analysis rules or on expert judgment, both of which struggle to generalize across projects and languages. This manuscript presents a machine-learning (ML) approach that uses features derived from static analysis—object-oriented metrics, control-flow measures, dependency signals, and lightweight lexical cues—to detect prominent smells such as God Class, Long Method, Feature Envy, Data Class, and Shotgun Surgery. We design a reproducible pipeline covering dataset construction, feature extraction, imbalance handling, model training, evaluation, and statistical testing. A simulation study emulates multi-project conditions and cross-version drift to approximate realistic industrial scenarios. Four supervised learners—Logistic Regression, Random Forest, SVM (RBF), and XGBoost—are compared under stratified cross-validation and cross-project holdout. Performance is reported using F1-score (primary), with secondary examinations of calibration, error structure, and explainability (via model-agnostic feature attribution).

Results show tree-based ensembles (Random Forest and XGBoost) consistently outperform linear and kernel baselines, particularly for class-imbalance-sensitive smells (e.g., Shotgun Surgery). Statistical analysis using non-parametric tests indicates significant differences among learners, and ablation suggests that combining structure-aware metrics with succinct lexical signals yields the best trade-off between accuracy and interpretability. We conclude with practical guidance for toolsmiths and teams: use ensemble ML as an assistive layer on top of static analysis, expose explainable rankings rather than hard flags, calibrate thresholds by smell type, and validate models cross-project to avoid overfitting to local coding styles.