Contextualised Actionable Explications ((12) A06*)

实现数据驱动的AI系统清晰度的关键一步是提高我们对这种系统为什么会这样做的理解。在这方面，反事实解释是有希望的。然而，它们可能无法使人类利益相关者根据得到的解释采取行动，也无法解决利益相关者的上下文需求。在项目A6中，我们的目标是从反事实的解释转向情境化的可操作解释。我们为机器学习算法开发了可操作解释的数学基础：揭示人类可行的决策相关原因的解释。在实证研究中，我们测试和改进我们的方法，从不同利益相关者的角度得到的解释的有用性。

A crucial step towards perspicuity of data-driven AI systems is to improve our understanding of why such a system acts the way it does. In this regard, counterfactual explications are promising. Yet, they may not enable human stakeholders to act upon resulting explanations, nor do they address stakeholders' contextual needs. In Project A6 we aim to move from counterfactual explications to contextualised actionable explications. We develop the mathematical foundations of actionable explications for machine learning algorithms: explications that uncover decision-relevant reasons that are feasible for humans to act upon. In empirical studies, we test and improve our approach with respect to the usefulness of resulting explanations from the perspective of different stakeholders.