Explaining predictive models using Shapley values and non-parametric vine copulas

Abstract In this paper the goal is to explain predictions from complex machine learning models. One method that has become very popular during last few years Shapley values. The original development of values for prediction explanation relied on assumption features being described were independent. If in reality are dependent may lead incorrect explanations. Hence, there have recently been attempts appropriately modelling/estimating dependence between features. Although previously proposed methods clearly outperform traditional approach assuming independence, they their weaknesses. we propose two new approaches modelling Both based vine copulas, which flexible tools multivariate non-Gaussian distributions able characterise a wide range dependencies. performance evaluated simulated data sets and real set. experiments demonstrate copula give more accurate approximations true than competitors.