Unifying local and global model explanations by functional decomposition of low dimensional structures

Conditional Expectation Network for SHAP

“…In the examples in the next section, we compare the conditional and unconditional versions (4.5) and (4.6), respectively, and for the extrapolation, we simply use the one provided by the fitted neural network. We remark that there is interesting work that extends Shapley values to higher order decompositions and representations; we refer to Tsai et al [34] and Hiabu et al [15]. The basic idea is to give a functional decomposition of the regression function by including higher interaction terms.…”

“…The basic idea is to give a functional decomposition of the regression function by including higher interaction terms. This can partly mitigate the difficulty of the decision whether one should work with conditional or unconditional expectations, however, some issues remain, e.g., the above mentioned support constraints cannot be dealt with the (unconstrained) marginal identification given by formula (2) in Hiabu et al [15].…”