In this pull request, we add support for including paired combination sampling as an option when using a sampled subset of combinations to compute the Shapley value explanations. The user can specify whether he/she wants to use unique (default before, and still default), unique-paired (new), or non-unique (default in very old version of shapr).
Paired sampling has been shown to improve the precision of the approximated Shapley value explanations in several papers., see e.g., (https://arxiv.org/abs/2107.07436).
In my own simulation studies, I found it beneficial to use paired sampling when the features were correlated, but I saw no improvement for independent features.
In this pull request, we add support for including paired combination sampling as an option when using a sampled subset of combinations to compute the Shapley value explanations. The user can specify whether he/she wants to use
unique
(default before, and still default),unique-paired
(new), ornon-unique
(default in very old version ofshapr
).Paired sampling has been shown to improve the precision of the approximated Shapley value explanations in several papers., see e.g., (https://arxiv.org/abs/2107.07436).
In my own simulation studies, I found it beneficial to use paired sampling when the features were correlated, but I saw no improvement for independent features.