Shapley Additive Explanations for Knowledge Discovery via Surrogate Models

Abstract

It is sometimes desirable to delve further into how the inputs affect the output in design optimization and uncertainty analysis. Surrogate models such as Gaussian Process Regression and support vector regression are useful for such tasks and can be further enhanced by introducing advanced post-processing methods. This paper investigates Shapley Additive Explanation (SHAP) as a tool to aid surrogate-assisted data-driven analysis. In particular, surrogate-enabled SHAP analysis allows visualization of the input-output relationship in a meaningful way using summary SHAP plots and SHAP dependence plots. Some important information that can be extracted and visualized from SHAP includes the importance of input variables (i.e., global sensitivity analysis), nonlinearity level, and level of interactions. The benefits of Shapley values for engineering analysis using surrogate models are demonstrated in three engineering test problems.