Empirical Stability Assessment of SHAP-Based Feature Selection Across Controlled Model Complexities

This study empirically assesses the stability and reliability of SHAP-based feature selection across varying model complexities by introducing a scalable custom model $f(x)$ and validating the approach using the CICIDS2017 dataset to address gaps in understanding how complexity influences feature attribution in black-box models.

Original authors: Cynthia C, Gopal Krishna Kamath, Debayani Ghosh, Prashant K. Wali