Evaluating Explanations: How much do explanations from the teacher aid students?
This work addresses the problem of evaluating explanation methods for the machine learning community, providing a more robust and less gameable evaluation framework.
This paper introduces a framework to quantify the value of explanations by measuring the accuracy gains they provide to a student model during training, where explanations are unavailable at test time. They compare various attribution methods for text classification and question answering, finding consistent quantitative differences across student model architectures and learning strategies.
While many methods purport to explain predictions by highlighting salient features, what aims these explanations serve and how they ought to be evaluated often go unstated. In this work, we introduce a framework to quantify the value of explanations via the accuracy gains that they confer on a student model trained to simulate a teacher model. Crucially, the explanations are available to the student during training, but are not available at test time. Compared to prior proposals, our approach is less easily gamed, enabling principled, automatic, model-agnostic evaluation of attributions. Using our framework, we compare numerous attribution methods for text classification and question answering, and observe quantitative differences that are consistent (to a moderate to high degree) across different student model architectures and learning strategies.