How to handle categorical features

[gmgeorg]

AFAICT the method works only for numeric features [that are scaled to [0, 1]^d] -- at least in paper.

For categorical features, what do you propose? Mean encode it first and then apply IGCS? or does IGCS allow a direct way to compute IGCS for categorical data (naively: categorical should be easier since they should lead to a summation to begin with, not integrals that have to be approximated -- but might be completely off here).

[naofumihama]

Thank you for your comments. Since IGCS only refers the similarity of features, when the categorical feature is encoded to numerical and its level is few enough, IGCS works directly by distinguishing them and picking the exact matching data as similar or not. However, note that there are examples where the results can differ depending on before/after one-hot encoding in Cohort Shapley case as shown in Section 7.1 in https://arxiv.org/abs/2205.15750 . IGCS take an integral on indicator space (coordinated by z in our paper), always continuum variable, so even if the feature x is discrete, the integral does not turn out to be a summation.