. Compared to the baseline γ = 0

Clearly, this does not provide a meaningful relaxation of the categorical constraint. We closely follow Fischer et al. [15]. With these variables, each term can be directly encoded as it consists of a linear function. In this section, we provide a detailed overview of the datasets considered in Section 6. Adult, German, Health, and Law School, have a highly skewed distribution of positive labels. Note, that the percentages do not sum to 100% as the labels are aggregated by patient and year.