Shapley values is one of the most popular model-agnostic feature attribution explanation methods. In this talk, first the theory of Shapley value is briefly introduced, especially, what are the desirable properties of an attribution explanation method and Shapley value is the only possible attribution explanation that satisfies most of the properties. Next, several algorithms that is able to exactly or approximately compute the Shapley value are discussed. Finally, the application of Shapley value in an ongoing question answering task is demonstrated.
Duration: 15min
Speaker: Zihao Wang