Evaluating explanation usefulness and fidelity

Overview

Separates whether an explanation satisfies stakeholders from whether it faithfully represents the model. It compares application-grounded, human-grounded and proxy-task evaluation and discusses trade-offs between interpretability and other model properties.