Abstract
Ranking functions play a crucial role in supporting decision-making processes across various critical domains. Given their widespread use, coupled with the fact that these functions are often directly learned from data, it is becoming more and more important to provide explanations that make the underlying models more transparent. In this paper, we propose the first formal approach to explain ranking functions. Our approach is model-agnostic, requiring only black-box access to the ranking function. We study the formal properties of this new approach, including an analysis of the complexity of computing an explanation. To demonstrate its feasibility, we implement our approach and conduct an experimental evaluation using as a case study a neural network model for predicting breast cancer recurrence.
