Merge pull request #42 from mayer79/release_candidate

Release candidate

CRAN-SUBMISSION

@@ -0,0 +1,3 @@
+Version: 0.3.0
+Date: 2022-09-29 15:39:13 UTC
+SHA: 0652b701dc44c6446c9ed4c92e4da3d76089f7b6

DESCRIPTION

@@ -1,22 +1,22 @@
 Package: kernelshap
 Title: Kernel SHAP
-Version: 0.2.0.900
+Version: 0.3.0
 Authors@R: c(
     person("Michael", "Mayer", , "[email protected]", role = c("aut", "cre")),
     person("David", "Watson", , "[email protected]", role = "ctb")
   )
 Description: Multidimensional refinement of the Kernel SHAP algorithm
     described in Ian Covert and Su-In Lee (2021)
-    <http://proceedings.mlr.press/v130/covert21a>.  Depending on the
-    number of features, Kernel SHAP values can be calculated exactly, by
-    sampling, or by a combination of the two. As soon as sampling is
-    involved, the algorithm iterates until convergence, and standard
-    errors are provided.  The package allows to work with any model that
+    <http://proceedings.mlr.press/v130/covert21a>.  The package allows to
+    calculate Kernel SHAP values in an exact way, by iterative sampling
+    (as in the reference above), or by a hybrid of the two.  As soon as
+    sampling is involved, the algorithm iterates until convergence, and
+    standard errors are provided.  The package works with any model that
     provides numeric predictions of dimension one or higher.  Examples
-    include linear regression, logistic regression (logit or probability
-    scale), other generalized linear models, generalized additive models,
-    and neural networks.  The package plays well together with
-    meta-learning packages like 'tidymodels', 'caret' or 'mlr3'.
+    include linear regression, logistic regression (on logit or
+    probability scale), other generalized linear models, generalized
+    additive models, and neural networks.  The package plays well together
+    with meta-learning packages like 'tidymodels', 'caret' or 'mlr3'.
     Visualizations can be done using the R package 'shapviz'.
 License: GPL (>= 2)
 Depends: 

NEWS.md

@@ -1,4 +1,4 @@
-# kernelshap 0.2.0.900 DEVEL
+# kernelshap 0.3.0
 
 ## Major improvements
 

R/kernelshap.R

@@ -5,7 +5,7 @@
 #' The function allows to calculate Kernel SHAP values in an exact way, by iterative sampling 
 #' as in CL21, or by a hybrid of these two options. As soon as sampling is involved, 
 #' the algorithm iterates until convergence, and standard errors are provided.
-#' The default behaviour depends on the number of features p:
+#' The default behaviour depends on the number of features p, see also Details below:
 #' \itemize{
 #'   \item 2 <= p <= 8: Exact Kernel SHAP values are returned (for the given background data). 
 #'   \item p > 8: Hybrid (partly exact) iterative version of Kernel SHAP
@@ -191,7 +191,7 @@ kernelshap.default <- function(object, X, bg_X, pred_fun = stats::predict, bg_w
     all(nms %in% colnames(bg_X)),
     is.function(pred_fun),
     exact %in% c(TRUE, FALSE),
-    p == 1L || hybrid_degree %in% 0:(p / 2),
+    p == 1L || exact || hybrid_degree %in% 0:(p / 2),
     paired_sampling %in% c(TRUE, FALSE),
     "m must be even" = trunc(m / 2) == m / 2
   )

README.md

@@ -4,7 +4,7 @@
 
 SHAP values (Lundberg and Lee, 2017) decompose model predictions into additive contributions of the features in a fair way. A model agnostic approach is called Kernel SHAP, introduced in Lundberg and Lee (2017), and investigated in detail in Covert and Lee (2021). 
 
-The "kernelshap" package implements a multidimensional refinement of the Kernel SHAP Algorithm described in Covert and Lee (2021). The package allows to calculate Kernel SHAP values in an exact way, by iterative sampling (as in Covert and Lee, 2021), or a hybrid of the two. As soon as sampling is involved, the algorithm iterates until convergence, and standard errors are provided.
+The "kernelshap" package implements a multidimensional refinement of the Kernel SHAP Algorithm described in Covert and Lee (2021). The package allows to calculate Kernel SHAP values in an exact way, by iterative sampling (as in Covert and Lee, 2021), or by a hybrid of the two. As soon as sampling is involved, the algorithm iterates until convergence, and standard errors are provided.
 
 The default behaviour depends on the number of features $p$:
 
@@ -283,7 +283,7 @@ fit <- gam(Sepal.Length ~ s(Sepal.Width) + Species, data = iris)
 
 system.time(
   s <- kernelshap(
-    fit, 
+    fit,  
     iris[c(2, 5)], 
     bg_X = iris, 
     parallel = TRUE, 
@@ -300,7 +300,7 @@ SHAP values of first 2 observations:
 
 ## Exact/sampling/hybrid
 
-In above examples, since $p$ was small, exact Kernel SHAP values were calculated. Here, we want to show how to use the different strategies (exact, hybrid, and pure sampling) in a situation with ten features. 
+In above examples, since $p$ was small, exact Kernel SHAP values were calculated. Here, we want to show how to use the different strategies (exact, hybrid, and pure sampling) in a situation with ten features, see `?kernelshap` for details about those strategies.
 
 With ten features, a degree 2 hybrid is being used by default: 
 
@@ -343,7 +343,7 @@ s$S[1:5]
 
 The results are identical. While more on-off vectors $z$ were required (1022), only a single call to `predict()` was necessary.
 
-Pure sampling can be enforced by setting the hybrid degree to 0:
+Pure sampling (not recommended!) can be enforced by setting the hybrid degree to 0:
 
 ```r
 s <- kernelshap(fit, X[1L, ], bg_X = X, hybrid_degree = 0)

compare_with_python.R

@@ -65,8 +65,8 @@ fit <- lm(
 X_small <- diamonds[seq(1, nrow(diamonds), 53), setdiff(names(diamonds), "price")]
 
 # Exact KernelSHAP on X_small, using X_small as background data 
-# (71/59 seconds for exact, 27/17 for hybrid deg 2, 17/9 for hybrid deg 1, 
-# 26/15 for pure sampling; second number with 2 parallel sessions on Windows)
+# (58/67(?) seconds for exact, 25/18 for hybrid deg 2, 16/9 for hybrid deg 1, 
+# 26/17 for pure sampling; second number with 2 parallel sessions on Windows)
 system.time(
   ks <- kernelshap(fit, X_small, bg_X = bg_X)  
 )

cran-comments.md

@@ -0,0 +1,32 @@
+Hello CRAN
+
+This is an update with
+
+- a much better way to calculate *exact* KernelSHAP values,
+- and a very fast and accurate hybrid between exact and sampling.
+
+Furthermore, some defaults have been improved. As the package is maturing, the next
+update will hopefully be version 1.0.0.
+
+## Checks
+
+### `check(manual = TRUE, cran = TRUE)`
+
+0 errors ✔ | 0 warnings ✔ | 0 notes ✔
+
+### `check_win_devel()`
+
+* checking for detritus in the temp directory ... NOTE
+Found the following files/directories:
+  'lastMiKTeXException'
+
+0 errors ✔ | 0 warnings ✔ | 1 note ✖
+
+### `check_rhub()`
+
+- Ubuntu Linux 20.04.1 LTS, R-release, GCC: Okay
+- Platform:	Fedora Linux, R-devel, clang, gfortran: Note
+
+* checking HTML version of manual ... NOTE
+Skipping checking HTML validation: no command 'tidy' found
+

packaging.R

@@ -15,14 +15,15 @@ library(usethis)
 use_description(
   fields = list(
     Title = "Kernel SHAP",
-    Version = "0.2.0.900",
+    Version = "0.3.0",
     Description = "Multidimensional refinement of the Kernel SHAP algorithm described in
     Ian Covert and Su-In Lee (2021) <http://proceedings.mlr.press/v130/covert21a>.
-    Depending on the number of features, Kernel SHAP values can be calculated exactly,
-    by sampling, or by a combination of the two. As soon as sampling is involved, 
-    the algorithm iterates until convergence, and standard errors are provided.
-    The package allows to work with any model that provides numeric predictions of dimension one or higher.
-    Examples include linear regression, logistic regression (logit or probability scale),
+    The package allows to calculate Kernel SHAP values in an exact way, by iterative 
+    sampling (as in the reference above), or by a hybrid of the two. 
+    As soon as sampling is involved, the algorithm iterates until convergence, 
+    and standard errors are provided.
+    The package works with any model that provides numeric predictions of dimension one or higher.
+    Examples include linear regression, logistic regression (on logit or probability scale),
     other generalized linear models, generalized additive models, and neural networks. 
     The package plays well together with meta-learning packages like 'tidymodels', 'caret' or 'mlr3'. 
     Visualizations can be done using the R package 'shapviz'.",