Appendix for "The Ground Truth Effect - SZZ variations" paper

This study investigates the impact of ground truth construction on the performance of just-in-time vulnerability prediction models. Specifically, it compares eight SZZ algorithm variants used to identify vulnerability-contributing commits (VCCs) in open-source Java projects. The analysis evaluates both the overlap among labeling techniques and their effect on model accuracy. Results show that differences in ground truth definitions can significantly influence model performance, especially for tree-based and boosting classifiers. Variants such as B-SZZ, V-SZZ, MA-SZZ, and VCC-SZZ lead to more stable and effective models compared to L-SZZ and R-SZZ.