{"id":1338,"date":"2023-11-11T10:59:02","date_gmt":"2023-11-11T09:59:02","guid":{"rendered":"https:\/\/lorentzen.ch\/?p=1338"},"modified":"2023-11-20T20:48:16","modified_gmt":"2023-11-20T19:48:16","slug":"permutation-shap-versus-kernel-shap","status":"publish","type":"post","link":"https:\/\/lorentzen.ch\/index.php\/2023\/11\/11\/permutation-shap-versus-kernel-shap\/","title":{"rendered":"Permutation SHAP versus Kernel SHAP"},"content":{"rendered":"\n

SHAP is the predominant way to interpret black-box ML models, especially for tree-based models with the blazingly fast TreeSHAP algorithm.<\/p>\n\n\n\n

For general models, two slower SHAP algorithms exist:<\/p>\n\n\n\n

    \n
  1. Permutation SHAP (\u0160trumbelj and Kononenko, 2010)<\/li>\n\n\n\n
  2. Kernel SHAP (Lundberg and Lee, 2017)<\/li>\n<\/ol>\n\n\n\n

    Kernel SHAP was introduced in [2] as an approximation to permutation SHAP. <\/p>\n\n\n\n

    The 0.4.0 CRAN release<\/a> of our {kernelshap} package now contains an exact permutation SHAP algorithm for up to 14 features, and thus it becomes easy to make experiments between the two approaches.<\/p>\n\n\n\n

    Some initial statements about permutation SHAP and Kernel SHAP<\/h3>\n\n\n\n
      \n
    1. Exact permutation SHAP and exact Kernel SHAP have the same computational complexity.<\/li>\n\n\n\n
    2. Technically, exact Kernel SHAP is still an approximation of exact permutation SHAP, so you should prefer the latter.<\/li>\n\n\n\n
    3. Kernel SHAP assumes feature independence. Since features are never independent in practice: does this mean we should never use Kernel SHAP?<\/li>\n\n\n\n
    4. Kernel SHAP can be calculated almost exactly for any number of features, while permutation SHAP approximations get more and more inprecise when the number of features gets too large.<\/li>\n<\/ol>\n\n\n\n

      Simulation 1<\/h3>\n\n\n\n

      We will first work with the iris data because it has extremely strong correlations between features. To see the impact of having models with and without interactions, we work with a random forest model of increasing tree depth. Depth 1 means no interactions, depth 2 means pairwise interactions etc.<\/p>\n\n\n\n

      library(kernelshap)\nlibrary(ranger)\n\ndifferences <- numeric(4)\n\nset.seed(1)\n\nfor (depth in 1:4) {\n  fit <- ranger(\n    Sepal.Length ~ ., \n    mtry = 3,\n    data = iris, \n    max.depth = depth\n  )\n  ps <- permshap(fit, iris[2:5], bg_X = iris)\n  ks <- kernelshap(fit, iris[2:5], bg_X = iris)\n  differences[depth] <- mean(abs(ks$S - ps$S))\n}\n\ndifferences  # for tree depth 1, 2, 3, 4\n# 5.053249e-17 9.046443e-17 2.387905e-04 4.403375e-04\n\n# SHAP values of first two rows with tree depth 4\nps\n#      Sepal.Width Petal.Length Petal.Width      Species\n# [1,]  0.11377616   -0.7130647  -0.1956012 -0.004437022\n# [2,] -0.06852539   -0.7596562  -0.2259017 -0.006575266\n  \nks\n#      Sepal.Width Petal.Length Petal.Width      Species\n# [1,]  0.11463191   -0.7125194  -0.1951810 -0.006258208\n# [2,] -0.06828866   -0.7597391  -0.2259833 -0.006647530<\/pre><\/div>\n\n\n\n