They help with interpretability, which is a step brother. See for example "From Shapley Values to Generalized Additive Models and back" https://arxiv.org/abs/2209.04012