Ensembles of Random SHAPs

Abstract

The ensemble-based modifications of the well-known SHapley Additive exPlanations (SHAP) method for the local explanation of a black-box model are proposed. The modifications aim to simplify the SHAP which is computationally expensive when there is a large number of features. The main idea behind the proposed modifications is to approximate the SHAP by an ensemble of SHAPs with a smaller number of features. According to the first modification, called the ER-SHAP, several features are randomly selected many times from the feature set, and the Shapley values for the features are computed by means of “small” SHAPs. The explanation results are averaged to obtain the final Shapley values. According to the second modification, called the ERW-SHAP, several points are generated around the explained instance for diversity purposes, and the results of their explanation are combined with weights depending on the distances between the points and the explained instance. The third modification, called the ER-SHAP-RF, uses the random forest for a preliminary explanation of the instances and determines a feature probability distribution which is applied to the selection of the features in the ensemble-based procedure of the ER-SHAP. Many numerical experiments illustrating the proposed modifications demonstrate their efficiency and properties for a local explanation.

Special Issue on Ensemble Learning and/or Explainability

“…Finally, the eighth paper is entitled "Ensembles of random SHAPs" and it is authored by Utkin and Konstantinov [11]. In this work, the authors proposed three ensemblebased modifications to the SHapley Additive exPlanations (SHAP) method for the local explanation of a black-box model, called ER-SHAP, ERW-SHAP, and ER-SHAP-RF.…”

Interpretable Ensembles of Hyper-Rectangles as Base Models

“…Various modifications of SHAP have been developed to explain different machine learning models and tools [46,47,48,49,50,51,52]. Applications of SHAP can be found in [53,54,55], Approaches to reduce the computational complexity of SHAP were also proposed in [56,57,58,59,60]. Many interpretation methods and their comparison were considered and studied in survey papers [61,62,63,64,65,66,67,68] in detail.…”

“…Value φ is calculated from the efficiency property as Thus, we need first to find values for all features for which the explained example is inside the rectangle using (58), and then using (60) to calculate Shapley values for the remaining features. In order to obtain Shapley values for an arbitrary HRBM ensemble, for example, boosting, we use the third property (linearity) of Shapley values.…”