A review of the transition from Shapley values and SHAP values to RGE

The paper reviews the transition from Shapley values Shapley LS. A value for n-person games. 1953;2 and SHapley Additive exPlanations (SHAP) Lundberg SM, Lee S. A unified approach to interpreting model predictions. In: Proceedings of NIPS2017; 2017. p 4768–4777 to Rank Graduation Explainability (RGE) Babaei G, Giudici P, Raffinetti E. A rank graduation box for safe AI. Expert Syst Appl. 2025;259:125239, exploring how these methods enhance explainability in machine learning (ML) and artificial intelligence (AI) models. Shapley values, initially used for the fair allocation of payoffs in cooperative game theory, are extended into SHAP values to evaluate feature importance, finding extensive applications in credit evaluation, economic analysis, neural networks and multiple other fields. However, the computational complexity associated with calculating Shapley values and SHAP values on high-dimensional datasets limits their practical utility in real-time or large-scale application scenarios. To address these challenges, RGE is proposed as a new metric for predictive explainability, providing a more concise and efficient method for assessing model performance by comparing full models with reduced models lacking specific variables. Recent studies which are about SHAP values also introduce improvements such as CF-SHAP, FF-SHAP, aiming to increase computational efficiency and explanatory power. Future research could further investigate the balance between theoretical rigorness and practical applicability, aiming to develop more precise and easier understandable tools to analyse complex models.