PriorLabs/tabpfn-extensions

Hello everyone, After looking into which approximators would work best for our SHAP-value computations in #342, I'm now wondering whether there's a way to evaluate the quality of the imputers at our disposal and make an informed choice for our default based on this. My basic idea for this was the following (bear with me, I am really not a SHAP expert, so please do criticize). The value function we want to approximate with our imputer is $v(S)=E_q[f(X)|X_s=x_S]$, where $q=p(x_\bar{S})$ (marginal) or $q=p(x_\bar{S}|x_S)$ (conditional). If I understand it correctly, the choice of $q$ depends on which formulation of the underlying game we chose to compute the SHAP values, but for the general case, where we want to explain the model behavior vs. the data, the marginal formulation is preferable. $v(S)$ can be pretty well estimated by just marginalizing over the training set, i.e., just taking the mean over $f(x_S, x_{i, \bar{S}})$ for all $x_i$ in the training set, so that could serve as a p