Make model agnostic SHAP for H2O more visible.

Author: mayer79

.github/workflows/test-coverage.yaml

@@ -40,8 +40,6 @@ jobs:
               "shapviz\\.shapr",
               "shapviz\\.kernelshap",
               "shapviz\\.permshap",
-              "shapviz\\.H2ORegressionModel",
-              "shapviz\\.H2OBinomialModel",
               "shapviz\\.H2OModel",
               "\\.onLoad"
             )

NAMESPACE

@@ -23,9 +23,7 @@ S3method(print,mshapviz)
 S3method(print,shapviz)
 S3method(rbind,mshapviz)
 S3method(rbind,shapviz)
-S3method(shapviz,H2OBinomialModel)
 S3method(shapviz,H2OModel)
-S3method(shapviz,H2ORegressionModel)
 S3method(shapviz,default)
 S3method(shapviz,explain)
 S3method(shapviz,kernelshap)

NEWS.md

@@ -2,7 +2,8 @@
 
 ### Documentation
 
-- H2O random forests (regression and binary classification) are now supported as well (fast TreeSHAP) [#163](https://github.com/ModelOriented/shapviz/pull/163).
+- H2O now supports passing background data for model agnostic SHAP. This is now easier visible in {shapviz}, see https://github.com/h2oai/h2o-3/issues/16463.
+- H2O random forests (regression and binary classification) now support TreeSHAP as well [#163](https://github.com/ModelOriented/shapviz/pull/163).
 
 ### Compatibility
 

R/shapviz.R

@@ -5,7 +5,7 @@
 #' from a fitted model of type
 #' - XGBoost,
 #' - LightGBM, or
-#' - H2O (tree-based models).
+#' - H2O.
 #'
 #' Furthermore, [shapviz()] can digest the results of
 #' - `fastshap::explain()`,
@@ -454,28 +454,26 @@ shapviz.kernelshap <- function(
 }
 
 #' @describeIn shapviz
-#'   Creates a "shapviz" object from a (tree-based) H2O regression model.
-#' @export
-shapviz.H2ORegressionModel = function(
-    object, X_pred, X = as.data.frame(X_pred), collapse = NULL, ...
-  ) {
-  shapviz.H2OModel(object = object, X_pred = X_pred, X = X, collapse = collapse, ...)
-}
-
-#' @describeIn shapviz
-#'   Creates a "shapviz" object from a (tree-based) H2O binary classification model.
-#' @export
-shapviz.H2OBinomialModel = function(
-    object, X_pred, X = as.data.frame(X_pred), collapse = NULL, ...
-  ) {
-  shapviz.H2OModel(object = object, X_pred = X_pred, X = X, collapse = collapse, ...)
-}
-
-#' @describeIn shapviz
-#'   Creates a "shapviz" object from a (tree-based) H2O model (base class).
+#'   Creates a "shapviz" object from an H2O model.
+#' @param background_frame Background dataset for baseline SHAP or marginal SHAP.
+#'    Only for H2O models.
+#' @param output_space If model has link function, this argument controls whether the
+#'   SHAP values should be linearly (= approximately) transformed to the original scale
+#'   (if `TRUE`). The default is to return the values on link scale.
+#'   Only for H2O models.
+#' @param output_per_reference Switches between different algorithms, see
+#'   `?h2o::h2o.predict_contributions` for details.
+#'   Only for H2O models.
 #' @export
 shapviz.H2OModel = function(
-    object, X_pred, X = as.data.frame(X_pred), collapse = NULL, ...
+    object,
+    X_pred,
+    X = as.data.frame(X_pred),
+    collapse = NULL,
+    background_frame = NULL,
+    output_space = FALSE,
+    output_per_reference = FALSE,
+    ...
   ) {
   if (!requireNamespace("h2o", quietly = TRUE)) {
     stop("Package 'h2o' not installed")
@@ -488,7 +486,16 @@ shapviz.H2OModel = function(
   if (!inherits(X_pred, "H2OFrame")) {
     X_pred <- h2o::as.h2o(X_pred)
   }
-  S <- as.matrix(h2o::h2o.predict_contributions(object, newdata = X_pred, ...))
+  S <- as.matrix(
+    h2o::h2o.predict_contributions(
+      object,
+      newdata = X_pred,
+      background_frame = background_frame,
+      output_space = output_space,
+      output_per_reference = output_per_reference,
+      ...
+    )
+  )
   shapviz.matrix(
     object = S[, setdiff(colnames(S), "BiasTerm"), drop = FALSE],
     X = X,

README.md

@@ -23,7 +23,7 @@
 
 SHAP and feature values are stored in a "shapviz" object that is built from:
 
-1. Models that know how to calculate SHAP values: XGBoost, LightGBM, H2O (tree-based models).
+1. Models that know how to calculate SHAP values: XGBoost, LightGBM, and H2O.
 2. SHAP crunchers like {fastshap}, {kernelshap}, {treeshap}, {fastr}, and {DALEX}.
 3. SHAP matrix and corresponding feature values.
 

man/shapviz.Rd

@@ -35,7 +35,7 @@ In particular, the following plots are available:
 
 SHAP and feature values are stored in a "shapviz" object that is built from:
 
-1. Models that know how to calculate SHAP values: XGBoost, LightGBM, h2o (boosted trees).
+1. Models that know how to calculate SHAP values: XGBoost, LightGBM, and h2o.
 2. SHAP crunchers like {fastshap}, {kernelshap}, {treeshap}, {fastr}, and {DALEX}.
 3. SHAP matrix and corresponding feature values.
 
@@ -178,7 +178,7 @@ sv_dependence(shp, "Sepal.Width")
 
 ### H2O
 
-If you work with a boosted trees H2O model:
+H2O supports TreeSHAP for boosted trees and random forests. For other models, model agnostic method based on marginal expectations are used, requiring a background dataset.
 
 ```r
 library(shapviz)
@@ -187,10 +187,18 @@ library(h2o)
 h2o.init()
 
 iris2 <- as.h2o(iris)
-fit <- h2o.gbm(colnames(iris[-1]), "Sepal.Length", training_frame = iris2)
-shp <- shapviz(fit, X_pred = iris)
-sv_force(shp, row_id = 1)
-sv_dependence(shp, "Species")
+
+# Random forest
+fit_rf <- h2o.randomForest(colnames(iris[-1]), "Sepal.Length", training_frame = iris2)
+shp_rf <- shapviz(fit_rf, X_pred = iris)
+sv_force(shp_rf, row_id = 1)
+sv_dependence(shp_rf, "Species")
+
+# Linear model
+fit_lm <- h2o.glm(colnames(iris[-1]), "Sepal.Length", training_frame = iris2)
+shp_lm <- shapviz(fit_lm, X_pred = iris, background_frame = iris2)
+sv_force(shp_lm, row_id = 1)
+sv_dependence(shp_lm, "Species")
 ```
 
 ### treeshap