[1S4-GS-2] Machine learning:

[1S4-GS-2-03] A Comparative Analysis of SHAP and Its Alternative Methods Based on Kernel Functions

Kazuhiro Hiraki², Shinichi Ishihara, 〇Junnosuke Shino Shino¹ (1. Waseda University, 2. International Monetary Fund)

Keywords:SHAP, AFA (Additive Feature Attribution), LIME, Kernel, Game Theory

Hiraki, Ishihara, and Shino [3] derive a general expression for Additive Feature Attribution (AFA) using arbitrary kernels and propose several AFAs that satisfy the condition that perturbed samples closer to the observation under analysis are assigned larger weights. This study reviews discussions on SHAP and its alternative methods and investigates, through numerical examples and real-world data, how these methods differ in their factor decomposition patterns. Distinct differences among these AFA methods were observed in both the numerical example and the empirical analysis of commodity price data (specifically, gold prices).