Paul-Gauthier Noé

Shapley values are extensively used to measure the contribution of the inputs to a function’s output. However, originating in cooperative game theory, they are defined for real-valued functions only, limiting their application.
In statistics and machine learning, models can output elements in a multidimensional vector space. As an example, in a classification task, the output space can be the multidimensional probability simplex. In this case, Shapley compositions, which are basically Shapley vectors in the Euclidean Aitchison simplex, can be defined. In this seminar, I will present this use case and discuss ongoing work on generalising Shapley's framework to vector-valued functions.