NEWS.md

kernelshap 0.9.0

Bug fix

With input from Mario Wuethrich and Ian Covert and his repo, we have fixed a bug in how kernelshap() calculates Kernel weights.

• The differences caused by this are typically very small.
• Models with interactions of order up to two have been unaffected.
• Exact Kernel SHAP now provides identical results to exact permutation SHAP.

Fixed in #168, which also has received unit tests against Python's "shap".

API

• The argument feature_names can now also be used with matrix input (#166).

Speed and memory improvements

• permshap() and kernelshap() require about 10% less memory (#166).
• permshap() and kernelshap() are faster for data.frame input, and slightly slower for matrix input (#166).
• Additionally, permshap(, exact = TRUE) is faster by pre-calculating more elements used across rows #165

Documentation

• kernelshap() and permshap() currently yield a warning on random seed handling in parallel mode, thanks #163 for reporting. We have added a note in the function documentation that this warning can be ignored.

Internal changes

• Matrices holding on-off vectors are now consistently of type logical (#167).

kernelshap 0.8.0

Major improvement

permshap() has received a sampling version, which is useful if the number of features p is larger than 8. The algorithm iterates until the resulting values are sufficiently precise. Additionally, standard errors are provided (#152).

During each iteration, the algorithm cycles twice through a random permutation: It starts with all feature components "turned on" (i.e., taking them from the observation to be explained), then gradually turning off components according to the permutation (i.e., marginalizing them over the background data). When all components are turned off, the algorithm - one by one - turns the components back on, until all components are turned on again. This antithetic scheme allows to evaluate Shapley's formula 2p times with each permutation, using a total of 2p + 1 evaluations of marginal means.

For models with interactions up to order two, one can show that even a single iteration provides exact SHAP values (with respect to the given background dataset).

The Python implementation "shap" uses a similar approach, but without providing standard errors, and without early stopping. To mimic its behavior, we would need to set max_iter = p in R, and max_eval = (2*p+1)*p in Python.

For faster convergence, we use balanced permutations in the sense that p subsequent permutations each start with a different feature. Furthermore, the 2p on-off vectors with sum <=1 or >=p-1 are evaluated only once, similar to the degree 1 hybrid in [kernelshap()] (but covering less weight).

User visible changes

• In exact mode, kernelshap() does not return the following elements anymore: m (= 0), converged (all TRUE), n_iter (all 1), and SE (all values 0) (#153).
• In sampling mode of kernelshap(), above elements have been moved to the end of the output list (#153).
• Removed unpaired sampling in kernelshap() (#154).
• The stopping criterion in sampling mode of kernelshap() used a slightly too strict convergence rule. This has been relaxed in #156.

Documentation

• New DESCRIPTION file.
• Adapted docstrings to reflect above changes (#155)

Maintenance

• Improve code coverage (#156).

Bug fixes

• kernelshap() with max_iter = 1 will now work (#160).

kernelshap 0.7.1 (not on CRAN)

Documentation

• More compact README.
• Updated function description.

Maintenance

• Update code coverage version (#150).

kernelshap 0.7.0

This release is intended to be the last before stable version 1.0.0.

Major change

Passing a background dataset bg_X is now optional.

If the explanation data X is sufficiently large (>= 50 rows), bg_X is derived as a random sample of bg_n = 200 rows from X. If X has less than bg_n rows, then simply bg_X = X. If X has too few rows (< 50), you will have to pass an explicit bg_X.

Minor changes

• ranger() survival models now also work out-of-the-box without passing a tailored prediction function. Use the new argument survival = "chf" in kernelshap() and permshap() to distinguish cumulative hazards (default) and survival probabilities per time point.
• The resulting object of kernelshap() and permshap() now contain bg_X and bg_w used to calculate the SHAP values.

kernelshap 0.6.0

Major changes

• Factor-valued predictions are not supported anymore.

Maintenance

• Fix CRAN note about unavailable link to gam::gam().
• Added dependency to {MASS} for calculating Moore-Penrose generalized matrix inverse.

kernelshap 0.5.0

New features

New additive explainer additive_shap() that works for models fitted via

• lm(),
• glm(),
• mgcv::gam(),
• mgcv::bam(),
• gam::gam(),
• survival::coxph(),
• survival::survreg().

The explainer uses predict(..., type = "terms"), a beautiful trick used in fastshap::explain.lm(). The result will be identical to those returned by kernelshap() and permshap() but exponentially faster. Thanks David Watson for the great idea discussed in #130.

User visible changes

• permshap() now returns an object of class "kernelshap" to reduce the number of redundant methods.
• To distinguish which algorithm has generated the "kernelshap" object, the outputs of kernelshap(), permshap() (and additive_shap()) got an element "algorithm".
• is.permshap() has been removed.

kernelshap 0.4.2

API

• {mlr3}: Non-probabilistic classification now works.
• {mlr3}: For probabilistic classification, you now have to pass predict_type = "prob".

Documentation

• The README has received an {mlr3} and {caret} example.

kernelshap 0.4.1

Performance improvements

• Significant speed-up for pure data.frames, i.e., no data.tables or tibbles.
• Some small performance improvements, e.g., for factor predictions and univariate predictions.
• Slight speed-up of permshap() by caching calculations for the two special permutations of all 0 and all 1. Consequently, the m_exact component in the output is reduced by 2.

Documentation

• Rewrote many examples in the README.
• Added reference to Erik Strumbelj and Ivan Kononeko (2014).

kernelshap 0.4.0

Major changes

• Added permshap() to calculate exact permutation SHAP values. The function currently works for up to 14 features.
• Factor-valued predictions are now supported. Each level is represented by its dummy variable.

Other changes

• Slight speed-up.
• Integer valued case weights are now turned into doubles to avoid integer overflow.

kernelshap 0.3.8

API improvements

• Multi-output case: column names of predictions are now used as list names of the resulting S and SE lists.

Bug fixes

• {mlr3} probabilistic classification would not work out-of-the-box. This has been fixed (with corresponding example in the README) in #100
• The progress bar was initialized at 1 instead of 0. This is fixed.

Maintenance

• Added explanation of sampling Kernel SHAP to help file.
• In internal calculations, use explicit feature_names as dimnames (#96).

kernelshap 0.3.7

Maintenance

• Fixed problem in Latex math for MacOS.

kernelshap 0.3.6

Maintenance

• Improved help files and README

kernelshap 0.3.5

Maintenance

• New contributor: Przemyslaw Biecek - welcome on board!
• My new cozy home: https://github.com/ModelOriented/kernelshap
• Webpage created with "pkgdown"
• Introduced Github workflows
• More unit tests

Small visible changes

• Removed the 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.
• Adding $(n \times K)$ matrix of predictions to the output, where $n$ is the number of rows in the explainer data X, and $K$ is the dimension of a single prediction (usually 1).
• Setting verbose = FALSE now does not suppress the warning on too large background data anymore. Use suppressWarnings() instead.

kernelshap 0.3.4

Documentation

• New logo
• Better package description
• Better README

kernelshap 0.3.3

Less dependencies

• Removed dependency "dorng". This might have an impact on the seeding if in parallel mode.
• Removed dependency "MASS"

kernelshap 0.3.2

Documentation

• Rewritten README and examples to better show the role of the background data.

Bug fixes

• When bg_X contained more columns than X, unflexible prediction functions could fail when being applied to bg_X.

kernelshap 0.3.1

Changes

• New argument 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.
• The background dataset can now consist of a single row only. This is useful in situations with natural "off" value such as for image data or for models that can naturally deal with missing values.

kernelshap 0.3.0

Major improvements

Exact calculations

Thanks to David Watson, exact calculations are now also possible for $p&gt;5$ 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 $p &gt; 10$.

Hybrid strategy

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:

1. Step 1 (exact part): There are $2p$ different on-off vectors $z$ with $\sum z \in {1, p-1}$, covering a large proportion of the Kernel SHAP distribution. The degree 1 hybrid will list those vectors and use them according to their weights in the upcoming calculations. Depending on $p$, we can also go a step further to a degree 2 hybrid by adding all $p(p-1)$ vectors with $\sum z \in {2, p-2}$ to the process etc. The necessary predictions are obtained along with other calculations similar to those in [1].
2. Step 2 (sampling part): The remaining weight is filled by sampling vectors $z$ according to Kernel SHAP weights renormalized to the values not yet covered by Step 1. Together with the results from Step 1 - correctly weighted - this now forms a complete iteration as in Covert and Lee (2021). The difference is that most mass is covered by exact calculations. Afterwards, the algorithm iterates until convergence. The output of Step 1 is reused in every iteration, leading to an extremely efficient strategy.

The default behaviour of kernelshap() is as follows:

• $p \le 8$: Exact Kernel SHAP (with respect to the background data)
• $9 \le p \le 16$: Degree 2 hybrid
• $p &gt; 16$: Degree 1 hybrid
• $p = 1$: Exact Shapley values

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.

User visible changes

• The default value of m is reduced from $8p$ to $2p$ except when hybrid_degree = 0 (pure sampling).
• The default value of exact is now TRUE for $p \le 8$ instead of $p \le 5$.
• A new argument 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).
• The default value of tol was reduced from 0.01 to 0.005.
• The default of max_iter was reduced from 250 to 100.
• The order of some of the arguments behind the first four has been changed.
• Paired sampling no longer duplicates m.
• Thanks to Mathias Ambuehl, the random sampling of z vectors is now fully vectorized.
• The output of print() is now more slim.
• A new summary() function shows more infos.

Other changes

• The resulting object now contains 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).

Bug fixes

• Predictions of mgcv::gam() would cause an error in check_pred() (they are 1D-arrays).
• Fixed small mistakes in the examples of the README (mlr3 and mgcv).

kernelshap 0.2.0

Breaking change

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:

• Logistic regression (logit scale): kernelshap(fit, X, bg_X)
• Logistic regression (probabilities): kernelshap(fit, X, bg_X, type = "response")
• Linear regression with logarithmic response, but evaluated on original scale: Here, the default predict function needs to be overwritten: kernelshap(fit, X, bg_X, pred_fun = function(m, X) exp(predict(m, X)))

Major improvements

• kernelshap() has received a more intuitive interface, see breaking change above.
• The package now supports multidimensional predictions. Hurray!
• Thanks to David Watson, parallel computing is now supported. The user needs to set up the parallel backend before calling 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().
• Even without parallel computing, kernelshap() has become much faster.
• For $2 \le p \le 5$ features, the algorithm now returns exact Kernel SHAP values with respect to the given background data. (For $p = 1$, exact Shapley values are returned.)
• Direct handling of "tidymodels" models.

User visible changes

• Besides 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).
• The default 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 $p &gt; 16$. The change will imply more robust results for large p.
• There were too many "ks_*()" functions to extract elements of a "kernelshap" object. They are now all deprecated and replaced by ks_extract(, what = "S").
• Added "MASS", "doRNG", and "foreach" to dependencies.

Bug fixes

• Depending on $m$ and $p$, the matrix inversion required in the constrained least-squares solution could fail. It is now replaced by MASS::ginv(), the Moore-Penrose pseudoinverse using svd().

New contributor

• David Watson

kernelshap 0.1.0

This is the initial release.