Understanding Unsupervised Learning Explanations Using Contextual Importance and Utility

While the concept of Explainability has advanced significantly in the past decade, many areas remain unexplored. Although XAI implementations have historically been employed in attempting to ‘open’ the traditional black-box model of supervised learning implementation aiming to extract human-understandable information, there are no successful attempts at tackling unsupervised learning. This paper aims to tackle the challenge of using an XAI approach, specifically Contextual Importance and Utility (CIU), in order to provide an explainability layer for unsupervised learning models. The paper introduces the current XAI approaches of CIU as well as the other state-of-the-art implementations such as Lime or Shapley. The challenges posed by the unsupervised learning problem are explored and discussed, both on a conceptual and technical level. A relatively novel approach using a CIU implementation on unsupervised clustering techniques is presented along with the brief comparison with another state-of-the-art method called LIME.