Explaining a probabilistic prediction on the simplex with Shapley compositions
Researchers have introduced a novel method called Shapley compositions to explain probabilistic predictions in multiclass machine learning models. This approach extends the traditional Shapley value concept, which is typically used for scalar predictions, to handle the complex compositional nature of multiclass probability distributions. By leveraging the Aitchison geometry from compositional data analysis, Shapley compositions offer a unique and axiomatic way to quantify feature contributions on the simplex, providing a more accurate explanation for multiclass predictions. AI
IMPACT Provides a more accurate method for interpreting multiclass AI model outputs, potentially improving trust and debugging.