Wasn’t Expecting that – Using Abnormality as a Key to Design a Novel User-Centric Explainable AI Method

Explainable artificial intelligence (XAI) aims at automatically generating user-centric explanations to help users scrutinize artificial intelligence (AI) decisions and establish trust in AI systems. XAI methods that generate counterfactual explanations are particularly promising as they mimic how humans construct explanations. Insights from the social sciences suggest that counterfactual explanations should convey the most abnormal causes that lead to the AI decision, as unexpected information fosters understanding. So far, no XAI method incorporates abnormality when generating counterfactual explanations. This paper aims to design a novel XAI method to generate abnormal counterfactual explanations. To this end, we propose a novel measure to quantify the abnormality of features in explanations and integrate it into a method to generate counterfactual explanations. We demonstrate the XAI method’s applicability on a real-world data set within the use case of house price prediction. We evaluate its efficacy following functionally-grounded and human-grounded evaluation. The results of our evaluation indicate that our method successfully integrates abnormality in generating counterfactual explanations. The resulting explanations are perceived as more helpful by users to scrutinize AI decisions and lead to higher trust in AI systems compared to state-of-the-art counterfactual explanations.