permshap() to calculate exact permutation SHAP values. The function currently works for up to 14 features.S and SE lists.feature_names as dimnames (https://github.com/ModelOriented/kernelshap/issues/96).ks_extract() function. It was designed to extract objects like the matrix S of SHAP values from the resulting “kernelshap” object x. We feel that the standard extraction options (x$S, x[["S"]], or getElement(x, "S")) are sufficient.X, and \(K\) is the dimension of a single prediction (usually 1).verbose = FALSE now does not suppress the warning on too large background data anymore. Use suppressWarnings() instead.bg_X contained more columns than X, unflexible prediction functions could fail when being applied to bg_X.feature_names allows to specify the features to calculate SHAP values for. The default equals to colnames(X). This should be changed only in situations when X (the dataset to be explained) contains non-feature columns.Thanks to David Watson, exact calculations are now also possible for \(p>5\) features. By default, the algorithm uses exact calculations for \(p \le 8\) and a hybrid strategy otherwise, see the next section. At the same time, the exact algorithm became much more efficient.
A word of caution: Exact calculations mean to create \(2^p-2\) on-off vectors \(z\) (cheap step) and evaluating the model on a whopping \((2^p-2)N\) rows, where \(N\) is the number of rows of the background data (expensive step). As this explodes with large \(p\), we do not recommend the exact strategy for \(p > 10\).
The iterative Kernel SHAP sampling algorithm of Covert and Lee (2021) [1] works by randomly sample \(m\) on-off vectors \(z\) so that their sum follows the SHAP Kernel weight distribution (renormalized to the range from \(1\) to \(p-1\)). Based on these vectors, many predictions are formed. Then, Kernel SHAP values are derived as the solution of a constrained linear regression, see [1] for details. This is done multiple times until convergence.
A drawback of this strategy is that many (at least 75%) of the \(z\) vectors will have \(\sum z \in \{1, p-1\}\), producing many duplicates. Similarly, at least 92% of the mass will be used for the \(p(p+1)\) possible vectors with \(\sum z \in \{1, 2, p-1, p-2\}\) etc. This inefficiency can be fixed by a hybrid strategy, combining exact calculations with sampling. The hybrid algorithm has two steps:
The default behaviour of kernelshap() is as follows:
It is also possible to use a pure sampling strategy, see Section “User visible changes” below. While this is usually not advisable compared to a hybrid approach, the options of kernelshap() allow to study different properties of Kernel SHAP and doing empirical research on the topic.
Kernel SHAP in the Python implementation “shap” uses a quite similar hybrid strategy, but without iterating. The new logic in the R package thus combines the efficiency of the Python implementation with the convergence monitoring of [1].
[1] Ian Covert and Su-In Lee. Improving KernelSHAP: Practical Shapley Value Estimation Using Linear Regression. Proceedings of The 24th International Conference on Artificial Intelligence and Statistics, PMLR 130:3457-3465, 2021.
m is reduced from \(8p\) to \(2p\) except when hybrid_degree = 0 (pure sampling).exact is now TRUE for \(p \le 8\) instead of \(p \le 5\).hybrid_degree is introduced to control the exact part of the hybrid algorithm. The default is 2 for \(4 \le p \le 16\) and degree 1 otherwise. Set to 0 to force a pure sampling strategy (not recommended but useful to demonstrate superiority of hybrid approaches).tol was reduced from 0.01 to 0.005.max_iter was reduced from 250 to 100.m.print() is now more slim.summary() function shows more infos.m_exact (the number of on-off vectors used for the exact part), prop_exact (proportion of mass treated in exact fashion), exact flag, and txt (the info message when starting the algorithm).mgcv::gam() would cause an error in check_pred() (they are 1D-arrays).The interface of kernelshap() has been revised. Instead of specifying a prediction function, it suffices now to pass the fitted model object. The default pred_fun is now stats::predict, which works in most cases. Some other cases are catched via model class (“ranger” and mlr3 “Learner”). The pred_fun can be overwritten by a function of the form function(object, X, ...). Additional arguments to the prediction function are passed via ... of kernelshap().
Some examples:
kernelshap(fit, X, bg_X)kernelshap(fit, X, bg_X, type = "response")kernelshap(fit, X, bg_X, pred_fun = function(m, X) exp(predict(m, X)))kernelshap() has received a more intuitive interface, see breaking change above.kernelshap(), e.g., using the “doFuture” package, and then set parallel = TRUE. Especially on Windows, sometimes not all global variables or packages are loaded in the parallel instances. These can be specified by parallel_args, a list of arguments passed to foreach().kernelshap() has become much faster.matrix, data.frames, and tibbles, the package now also accepts data.tables (if the prediction function can deal with them).kernelshap() is less picky regarding the output structure of pred_fun().kernelshap() is less picky about the column structure of the background data bg_X. It should simply contain the columns of X (but can have more or in different order). The old behaviour was to launch an error if colnames(X) != colnames(bg_X).m = "auto" has been changed from trunc(20 * sqrt(p)) to max(trunc(20 * sqrt(p)), 5 * p. This will have an effect for cases where the number of features \(p > 16\). The change will imply more robust results for large p.ks_extract(, what = "S").MASS::ginv(), the Moore-Penrose pseudoinverse using svd().This is the initial release.