To enhance plausibility, explanations should be class-discriminative, that is, they should differ depending on the target class. In image classification, for instance, the explanantia supporting a “car” label should highlight different regions than those supporting “dog”. This makes it easier for humans to understand the specific rationale for each class.
Class discriminativeness is commonly measured by comparing explanantia generated for different classes on the same input. Several comparison setups have been proposed: between the most and least likely classes [Li et al. (2020b), Rebuffi et al. (2020)], the top predicted vs. a randomly chosen class [Sixt et al. (2020)], or simply between the two classes in binary settings [Pope et al. (2019)].
Any suitable similarity or distance measure can be used to assess the degree of overlap between explanantia, such as the $L_0$ or $L_2$ distance [Nie et al. (2018), Pope et al. (2019)], rank correlation [Rebuffi et al. (2020)], or SSIM for image domain [Sixt et al. (2020)]. A lower similarity implies a clearer distinction between the rationales for each class.
Although the reported implementations target FA, the idea of class-discriminative explanations can be naturally extended to other type, wherever explanations can be generated for multiple class hypotheses.