Smell-Augmented Symbolic Regression for Explainable Cross-Project Defect Prediction

Cross-Project Defect Prediction (CPDP) remains a critical challenge in software engineering, primarily due to severe dataset shift and systemic class imbalance between source and target repositories. Furthermore, the state-of-the-art black-box machine learning models have an inherent lack of interpretability and out-of-distribution generalization. This paper presents an empirical framework leveraging Smell-Enhanced Symbolic Regression (SR) to derive explicit, human-readable mathematical formulations for defect prediction, capable of robust extrapolation. The proposed framework is rigorously evaluated across eight heterogeneous software project transitions against 23 baseline configurations using five standardized performance metrics and non-parametric significance analysis. The empirical results demonstrate that Symbolic Regression delivers complete interpretability without sacrificing the predictive performance of complex, top-tier models.