Research on Balancing Methods for Class Imbalance in Multi-Label Code Smell Detection

Code smell detection addresses software quality deterioration caused by poor implementations. Existing research mainly focuses on single-label detection, neglecting multi-label scenarios, and faces class imbalance challenges. This study simultaneously tackles multi-label detection and class imbalance using the MLCQ dataset with CK metrics-derived features. A dual-strategy framework combines problem transformation (e.g., Classifier Chain) and algorithm adaptation with Random Forest/CNN base models. Class imbalance is mitigated through hybrid sampling techniques and a novel function-fitted oversampling method alongside loss function adjustments. Experiments demonstrate the Classifier Chain's superiority in multi-label classification and hybrid sampling's effectiveness in balancing data distribution. Our method offers practical advancements for software quality assurance through improved detection precision.