Effect of using only 1024 combinations in Shapley Value analysis

Hi,

I calculated Shapley Values for an analysis I'm doing. My analysis has 30 predictors, which means I can have up to 2^30 combinations for calculating Shapley Values. Since using all combination is impossible (i.e. MATLAB just freezes), I used the default 1024 combinations, but I'm not sure how accurate my results are in this case.

Is there a way to directly assess the effect of using this relatively small number of combinations, or the relationship between the number of predictors and the number of coalitions? I only found a paper article on the Shapley uncertainty (here), but it is heavily mathematical and I couldn't use it at all.

Thanks,

[John D'Errico]

Since I have absolutely no expertise in this, I cannot say. (I assume you are using the shapely tool in MATLAB.) But if you are worried about how well it does for the default of 1024, then try it with some other choices, some larger, some smaller. Do you get qualitatively similar results? If not, then perhaps you should not trust the result from the default for this large of a problem.