Interaction Decomposition of prediction function

This paper discusses the foundation of methods for accurately grasping interaction effects. The partial dependence (PD) and accumulated local effects (ALE) methods, which capture interaction effects as terms, are known as global model-agnostic methods in the interpretable machine learning field. ALE provides a functional decomposition of the prediction function. In the present study, we propose and mathematically formalize the requirements of an interaction decomposition (ID) that decomposes a prediction function into its main and interaction effect terms. We also present a theorem by which a decomposition method meeting these requirements can be generated. Furthermore, we confirm that ALE is an ID but PD is not. Finally, we present examples of decomposition methods that meet the requirements of ID, using both existing methods and methods that differ from the existing ones.