Do Feature Attribution Methods Correctly Attribute Features?

Feature attribution methods are popular in interpretable machine learning. These compute the of each input feature to represent its importance, but there is no consensus on definition "attribution", leading many competing with little systematic evaluation, complicated particular by lack ground truth attribution. To address this, we propose a dataset modification procedure induce such truth. Using this procedure, evaluate three common methods: saliency maps, rationales, and attentions. We identify several deficiencies add new perspectives growing body evidence questioning correctness reliability these applied datasets wild. further discuss possible avenues for remedy recommend be tested against before deployment. The code appendix available at https://yilunzhou.github.io/feature-attribution-evaluation/.