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 version of the Kernel SHAP Algorithm 1 described in the supplement of Covert and Lee (2021). The behaviour depends on the number of features \(p\):
The main function kernelshap() has four key
arguments:
object: Fitted model object.X: A (n x p) matrix,
data.frame, tibble or data.table
of rows to be explained. Important: The columns should only represent
model features, not the response.bg_X: Background data used to integrate out “switched
off” features, often a subset of the training data (around 100 to 200
rows). It should contain the same columns as X. Columns not
in X are silently dropped and the columns are arranged into
the order as they appear in X.pred_fun: Prediction function of the form
function(object, X, ...), providing K >= 1 numeric
predictions per row. Its first argument represents the model
object, its second argument a data structure like
X. (The names of the first two arguments do not matter.)
Additional (named) arguments are passed via ... of
kernelshap(). The default, stats::predict,
will work in most cases. Some exceptions (classes “ranger” and mlr3
“Learner”) are handled separately. In other cases, the function must be
specified manually.Additional arguments of kernelshap() can be used to
control details of the algorithm and to activate parallel
processing.
Remarks
bg_w allows to weight background data according
to case weights.X is a matrix or
tibble.# install.packages("devtools")
devtools::install_github("mayer79/kernelshap")library(kernelshap)
library(shapviz)
fit <- lm(Sepal.Length ~ ., data = iris)
# Crunch SHAP values
s <- kernelshap(fit, iris[-1], bg_X = iris)
s
# Output (partly)
# SHAP values of first 2 observations:
# Sepal.Width Petal.Length Petal.Width Species
# [1,] 0.21951350 -1.955357 0.3149451 0.5823533
# [2,] -0.02843097 -1.955357 0.3149451 0.5823533
#
Corresponding standard errors:
Sepal.Width Petal.Length Petal.Width Species
[1,] 0 0 0 0
[2,] 0 0 0 0
# Plot with shapviz
shp <- shapviz(s)
sv_waterfall(shp, 1)
sv_importance(shp)
sv_dependence(shp, "Petal.Length")
library(kernelshap)
library(shapviz)
fit <- glm(
I(Species == "virginica") ~ Sepal.Length + Sepal.Width, data = iris, family = binomial
)
# Crunch SHAP values
s <- kernelshap(fit, iris[1:2], bg_X = iris, type = "response")
# Plot with shapviz
shp <- shapviz(s)
sv_waterfall(shp, 51)
sv_dependence(shp, "Sepal.Length")
library(ranger)
library(kernelshap)
set.seed(1)
fit <- ranger(Species ~ ., data = iris, probability = TRUE)
s <- kernelshap(fit, iris[c(1, 51, 101), -5], bg_X = iris)
s
# 'kernelshap' object representing
# - 3 SHAP matrices of dimension 3 x 4
# - feature data.frame/matrix of dimension 3 x 4
# - baseline: 0.3332606 0.3347014 0.332038
# - average iterations: 2
# - rows not converged: 0
#
# SHAP values of first 2 observations:
# [[1]]
# Sepal.Length Sepal.Width Petal.Length Petal.Width
# [1,] 0.02299032 0.008345772 0.3168792 0.3185242
# [2,] -0.01479279 -0.002066122 -0.1585125 -0.1578892
#
# [[2]]
# Sepal.Length Sepal.Width Petal.Length Petal.Width
# [1,] 0.00190803 -0.004574378 -0.1615047 -0.1705303
# [2,] 0.02443698 0.006684421 0.3157610 0.2942638
#
# [[3]]
# Sepal.Length Sepal.Width Petal.Length Petal.Width
# [1,] -0.024898347 -0.003771394 -0.1553745 -0.1479938
# [2,] -0.009644196 -0.004618300 -0.1572485 -0.1363747library(tidymodels)
library(kernelshap)
iris_recipe <- iris %>%
recipe(Sepal.Length ~ .)
reg <- linear_reg() %>%
set_engine("lm")
iris_wf <- workflow() %>%
add_recipe(iris_recipe) %>%
add_model(reg)
fit <- iris_wf %>%
fit(iris)
ks <- kernelshap(fit, iris[, -1], bg_X = iris)
kslibrary(mlr3)
library(mlr3learners)
library(kernelshap)
library(shapviz)
mlr_tasks$get("iris")
tsk("iris")
task_iris <- TaskRegr$new(id = "iris", backend = iris, target = "Sepal.Length")
fit_lm <- lrn("regr.lm")
fit_lm$train(task_iris)
s <- kernelshap(fit_lm, iris, bg_X = iris)
sv <- shapviz(s)
sv_dependence(sv, "Species")
library(caret)
library(kernelshap)
library(shapviz)
fit <- train(
Sepal.Length ~ .,
data = iris,
method = "lm",
tuneGrid = data.frame(intercept = TRUE),
trControl = trainControl(method = "none")
)
s <- kernelshap(fit, iris[1, -1], predict, bg_X = iris)
sv <- shapviz(s)
sv_waterfall(sv, 1)
library(kernelshap)
library(keras)
library(shapviz)
model <- keras_model_sequential()
model %>%
layer_dense(units = 6, activation = "tanh", input_shape = 3) %>%
layer_dense(units = 1)
model %>%
compile(loss = "mse", optimizer = optimizer_nadam(0.005))
model %>%
fit(
x = data.matrix(iris[2:4]),
y = iris[, 1],
epochs = 50,
batch_size = 30
)
X <- data.matrix(iris[2:4])
# Crunch SHAP values
system.time(
s <- kernelshap(model, X, bg_X = X, batch_size = 1000)
)
# Plot with shapviz (results depend on neural net seed)
shp <- shapviz(s)
sv_waterfall(shp, 1)
sv_importance(shp)
sv_dependence(shp, "Petal.Length")
As long as you have set up a parallel processing backend, parallel
computing is supported via foreach and %dorng.
The latter ensures that set.seed() will lead to
reproducible results.
library(kernelshap)
library(doFuture)
# Set up parallel backend
registerDoFuture()
plan(multisession, workers = 2) # Windows
# plan(multicore, workers = 2) # Linux, macOS, Solaris
fit <- stats::lm(Sepal.Length ~ ., data = iris)
# With parallel computing
system.time(
s <- kernelshap(fit, iris[, -1], bg_X = iris, parallel = TRUE)
)On Windows, sometimes not all packages or global objects are passed
to the parallel sessions. In this case, the necessary instructions to
foreach can be specified through a named list via
parallel_args, see the following example:
library(mgcv)
library(kernelshap)
library(doFuture)
# Set up parallel backend
registerDoFuture()
plan(multisession, workers = 2)
fit <- gam(Sepal.Length ~ s(Sepal.Width) + Species, data = iris)
s <- kernelshap(
fit, iris[-1], bg_X = iris, parallel = TRUE, parallel_args = list(.packages = "mgcv")
)
s
SHAP values of first 2 observations:
Sepal.Width Petal.Length Petal.Width Species
[1,] 0.35570963 -5.551115e-17 2.775558e-16 -1.135187
[2,] -0.04607082 3.552714e-15 3.885781e-15 -1.135187[1] Scott M. Lundberg and Su-In Lee. A Unified Approach to Interpreting Model Predictions. Advances in Neural Information Processing Systems 30, 2017.
[2] 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.