A Series of Unfortunate Counterfactual Events: the Role of Time in Counterfactual Explanations

Counterfactual explanations are a prominent example of post-hoc\ninterpretability methods in the explainable Artificial Intelligence research\ndomain. They provide individuals with alternative scenarios and a set of\nrecommendations to achieve a sought-after machine learning model outcome.\nRecently, the literature has identified desiderata of counterfactual\nexplanations, such as feasibility, actionability and sparsity that should\nsupport their applicability in real-world contexts. However, we show that the\nliterature has neglected the problem of the time dependency of counterfactual\nexplanations. We argue that, due to their time dependency and because of the\nprovision of recommendations, even feasible, actionable and sparse\ncounterfactual explanations may not be appropriate in real-world applications.\nThis is due to the possible emergence of what we call "unfortunate\ncounterfactual events." These events may occur due to the retraining of machine\nlearning models whose outcomes have to be explained via counterfactual\nexplanation. Series of unfortunate counterfactual events frustrate the efforts\nof those individuals who successfully implemented the recommendations of\ncounterfactual explanations. This negatively affects people's trust in the\nability of institutions to provide machine learning-supported decisions\nconsistently. We introduce an approach to address the problem of the emergence\nof unfortunate counterfactual events that makes use of histories of\ncounterfactual explanations. In the final part of the paper we propose an\nethical analysis of two distinct strategies to cope with the challenge of\nunfortunate counterfactual events. We show that they respond to an ethically\nresponsible imperative to preserve the trustworthiness of credit lending\norganizations, the decision models they employ, and the social-economic\nfunction of credit lending.\n