rcaretshap

SHAP (Shapley Additive Explanations) with caret in R

I have used the caret R package to train a neural network, and a random forest. Can I find the SHAP values for feature importance in any way?

Solution

  • Slightly modified from kernelshap's README: https://github.com/ModelOriented/kernelshap

    library(caret)
library(kernelshap)
library(shapviz)

fit <- train(
  Sepal.Length ~ . + Species * Sepal.Width, 
  data = iris, 
  method = "lm", 
  tuneGrid = data.frame(intercept = TRUE),
  trControl = trainControl(method = "none")
)

xvars <- colnames(iris[-1])
s <- kernelshap(fit, iris, predict, bg_X = iris, feature_names = xvars)
sv <- shapviz(s)
sv_importance(sv)
sv_dependence(sv, xvars)

    enter image description here enter image description here

    Remarks

    1. Replace the linear model by anything else
    2. If the dataset is larger than 500 obs, replace bg_X by a subsample of about 200--500 rows.
    3. Probabilistic classification will work similarly