How Explainable Is Explainability? Towards Better Metrics for Explainable AI

Despite the fact that machine learning has been applied in innumerable domains, its models have usually operated in a black box fashion, i.e. without revealing the rationale behind their decisions. For human users, insufficient model transparency may result in the lack of trust in the technology, effectively hindering its development and adoption. This pressing need for the society to understand the reasoning for the model’s decisions gave rise to the concept of explainable AI (xAI), which has been the subject of extensive research. Since its introduction, a number of techniques providing explainability have been proposed. Yet, there has been no consensus reached regarding how to measure and evaluate the performance of the explainability methods. So far, the state-of-the-art literature has proposed two directions for evaluating explainable AI–the technical one and the human-centered one. Although the literature deems the technical way to be the objective one, it still suggests supplementing the evaluation with the subjective, human-centered approach, which proves to be time-consuming and requires considerable effort. This paper highlights the need to enhance and improve the existing technical metrics, to quantify the explainability of ML models in an objective, automated and time and cost-effective way. The text presents the current state-of-the-art in xAI evaluation metrics, discussing both human-centered and computer-centered approaches. Its contribution is in its comprehensive discussion of the evaluation methods for explainable AI (xAI).