Explainability in Reservoir Well-logging Evaluation: Comparison of Variable Importance Analysis with Shapley Value Regression, SHAP and LIME
Shaogui Deng, Chris Aldrich, Xiu Liu, Fengjiao Zhang · IFAC-PapersOnLine · 2024
Machine learning algorithms have made significant progress in the logging evaluation field, but their "black box" characteristics make is a hindrance to the interpretation and acceptance of these models. In the face of a multitude of interpretation methods, choosing the most suitable becomes a challenge. In this study, a random forest regression model was first applied to a simulated dataset. Subsequently, Shapley value regression, two variants of Sh apley A dditive Ex p lanation (SHAP), i.e. KernelSHAP and TreeSHAP, and Local Interpretable Model-Agnostic Explanation (LIME) methods were compared for global and local explanation of the random forest model. In addition, a comparative analysis of these explanatory methods was carried out using shear wave velocity prediction as an example. It was found that all four methods provided an accurate analysis of the variable importance rankings. Regardless of the global or local interpretation, the Shapley value method gave the most reliable results, closely followed by the KernelSHAP method. In addition, it was observed that LIME outperformed TreeSHAP in terms of global interpretation, but not necessarily in terms of local explanation.