This metric applies to explanations that generate both a factual and counterfactual explanans, aiming to evaluate how well they reflect and contrast the model's reasoning.
Originally proposed by [Liu et al. (2021b)] for the graph domain, requiring both explanantia being subgraphs, it can be generalized to other input types.
The method computes the model's output for the factual and counterfactual explanantia and compares them to the original prediction using the negative symmetric Kullback-Leibler divergence. The difference between these two scores is normalized by the distance between the factual and counterfactual explanantia. A high normalized score indicates that both explanans are informative: the factual preserves the original reasoning, while the counterfactual shifts the decision appropriately.