Model interpretation of Inclusion-Exclusion integral neural network using SHAP Values and Owen Values

Abstract

Inclusion-Exclusion integral neural network is a network model that uses fuzzy measures, which are non-additive measures, and Inclusion-Exclusion integral, which are represented by polynomial operations. The model outputs of this network can be expressed in terms of Inclusion-Exclusion integral formula using the Möbius inversion formula, and the contribution of each feature to the prediction can be calculated using the Shapley values, which is one of the reward calculation methods used in the theory of cooperative games. This allows us to construct a network with interpretability for parameters after training. For this network, we interpret the model using the Shapley values and the Owen values, which is an extended version of the Shapley values when considering a particular coalition structure or division of the set of players. We will also examine model interpretation using SHAP values, where Shapley values are applied to machine learning to allow interpretation of the impact of features on prediction based on input-output relationships.