remove catboost support

ModelOriented · Sep 28, 2023 · fff85eb · fff85eb
1 parent cd54efd
commit fff85eb
Show file tree

Hide file tree

Showing 31 changed files with 60 additions and 567 deletions.
diff --git a/DESCRIPTION b/DESCRIPTION
@@ -14,7 +14,7 @@ Authors@R: c(
 Description: An efficient implementation of the TreeSHAP algorithm
     introduced by Lundberg et al., (2020) <doi:10.1038/s42256-019-0138-9>.
     It is capable of calculating SHAP values for tree-based models in
-    polynomial time.  Currently supported models include 'catboost',
+    polynomial time.  Currently supported models include 
     'gbm', 'randomForest', 'ranger', 'xgboost', 'lightgbm'.
 License: GPL-3
 URL: https://modeloriented.github.io/treeshap/,

diff --git a/NAMESPACE b/NAMESPACE
@@ -3,14 +3,12 @@
 S3method(predict,model_unified)
 S3method(print,model_unified)
 S3method(print,treeshap)
-S3method(unify,catboost.Model)
 S3method(unify,default)
 S3method(unify,gbm)
 S3method(unify,lgb.Booster)
 S3method(unify,randomForest)
 S3method(unify,ranger)
 S3method(unify,xgb.Booster)
-export(catboost.unify)
 export(colors_breakdown_drwhy)
 export(colors_discrete_drwhy)
 export(gbm.unify)

diff --git a/NEWS.md b/NEWS.md
@@ -1,6 +1,10 @@
-treeshap 0.2.1
+treeshap 0.2.2
 ----------------------------------------------------------------
 * Initial CRAN submission.
+* Removed `catboost.unify` function (as the `catboost` package is not on CRAN)
+
+treeshap 0.2.1
+----------------------------------------------------------------
 * Fixed `randomForest.unify` for classifiers ([#12](https://github.com/ModelOriented/treeshap/issues/12), [#23](https://github.com/ModelOriented/treeshap/issues/23))
 * Implemented consolidated (generic) `unify` function ([#18](https://github.com/ModelOriented/treeshap/issues/18))
 * An error is thrown when the data passed to the `unify` or `treeshap` functions contain variables that are not used by the model ([#14](https://github.com/ModelOriented/treeshap/issues/14))

diff --git a/R/model_unified.R b/R/model_unified.R
@@ -31,8 +31,6 @@
 #'
 #' \code{\link{gbm.unify}} for \code{\link[gbm:gbm]{GBM models}}
 #'
-#' \code{\link{catboost.unify}} for \code{\link[catboost:catboost.train]{Catboost models}}
-#'
 #' \code{\link{xgboost.unify}} for \code{\link[xgboost:xgboost]{XGBoost models}}
 #'
 #' \code{\link{ranger.unify}} for \code{\link[ranger:ranger]{ranger models}}

diff --git a/R/set_reference_dataset.R b/R/set_reference_dataset.R
@@ -19,8 +19,6 @@
 #'
 #' \code{\link{gbm.unify}} for \code{\link[gbm:gbm]{GBM models}}
 #'
-#' \code{\link{catboost.unify}} for \code{\link[catboost:catboost.train]{Catboost models}}
-#'
 #' \code{\link{xgboost.unify}} for \code{\link[xgboost:xgboost]{XGBoost models}}
 #'
 #' \code{\link{ranger.unify}} for \code{\link[ranger:ranger]{ranger models}}

diff --git a/R/treeshap.R b/R/treeshap.R
@@ -21,7 +21,6 @@
 #' \code{\link{xgboost.unify}} for \code{XGBoost models}
 #' \code{\link{lightgbm.unify}} for \code{LightGBM models}
 #' \code{\link{gbm.unify}} for \code{GBM models}
-#' \code{\link{catboost.unify}} for \code{catboost models}
 #' \code{\link{randomForest.unify}} for \code{randomForest models}
 #' \code{\link{ranger.unify}} for \code{ranger models}
 #' \code{\link{ranger_surv.unify}} for \code{ranger survival models}

diff --git a/R/unify.R b/R/unify.R
@@ -3,7 +3,7 @@
 #' Convert your tree-based model into a standardized representation.
 #' The returned representation is easy to be interpreted by the user and ready to be used as an argument in \code{treeshap()} function.
 #'
-#' @param model A tree-based model object of any supported class (\code{gbm}, \code{lgb.Booster}, \code{randomForest}, \code{ranger}, \code{xgb.Booster}, or \code{catboost.Model}).
+#' @param model A tree-based model object of any supported class (\code{gbm}, \code{lgb.Booster}, \code{randomForest}, \code{ranger}, or \code{xgb.Booster}).
 #' @param data Reference dataset. A \code{data.frame} or \code{matrix} with the same columns as in the training set of the model. Usually dataset used to train model.
 #' @param ... Additional parameters passed to the model-specific unification functions.
 #'
@@ -14,8 +14,6 @@
 #'
 #' \code{\link{gbm.unify}} for \code{\link[gbm:gbm]{GBM models}}
 #'
-#' \code{\link{catboost.unify}} for \code{\link[catboost:catboost.train]{CatBoost models}}
-#'
 #' \code{\link{xgboost.unify}} for \code{\link[xgboost:xgboost]{XGBoost models}}
 #'
 #' \code{\link{ranger.unify}} for \code{\link[ranger:ranger]{ranger models}}
@@ -73,11 +71,6 @@ unify.xgb.Booster <- function(model, data, recalculate = FALSE, ...){
   xgboost.unify(model, data, recalculate)
 }
 
-#' @export
-unify.catboost.Model <- function(model, data, recalculate = FALSE, ...){
-  catboost.unify(model, data, recalculate)
-}
-
 #' @export
 unify.default <- function(model, data, ...){
   stop("Provided model is not of type supported by treeshap.")

diff --git a/R/unify_catboost.R b/R/unify_catboost.R
diff --git a/R/unify_gbm.R b/R/unify_gbm.R
@@ -14,8 +14,6 @@
 #' @seealso
 #' \code{\link{lightgbm.unify}} for \code{\link[lightgbm:lightgbm]{LightGBM models}}
 #'
-#' \code{\link{catboost.unify}} for  \code{\link[catboost:catboost.train]{CatBoost models}}
-#'
 #' \code{\link{xgboost.unify}} for \code{\link[xgboost:xgboost]{XGBoost models}}
 #'
 #' \code{\link{ranger.unify}} for \code{\link[ranger:ranger]{ranger models}}

diff --git a/R/unify_lightgbm.R b/R/unify_lightgbm.R
@@ -18,8 +18,6 @@
 #'
 #' \code{\link{gbm.unify}} for \code{\link[gbm:gbm]{GBM models}}
 #'
-#' \code{\link{catboost.unify}} for  \code{\link[catboost:catboost.train]{CatBoost models}}
-#'
 #' \code{\link{xgboost.unify}} for \code{\link[xgboost:xgboost]{XGBoost models}}
 #'
 #' \code{\link{ranger.unify}} for \code{\link[ranger:ranger]{ranger models}}

diff --git a/R/unify_randomForest.R b/R/unify_randomForest.R
@@ -20,8 +20,6 @@
 #'
 #' \code{\link{gbm.unify}} for \code{\link[gbm:gbm]{GBM models}}
 #'
-#' \code{\link{catboost.unify}} for  \code{\link[catboost:catboost.train]{CatBoost models}}
-#'
 #' \code{\link{xgboost.unify}} for \code{\link[xgboost:xgboost]{XGBoost models}}
 #'
 #' \code{\link{ranger.unify}} for \code{\link[ranger:ranger]{ranger models}}

diff --git a/R/unify_ranger.R b/R/unify_ranger.R
@@ -18,8 +18,6 @@
 #'
 #' \code{\link{gbm.unify}} for \code{\link[gbm:gbm]{GBM models}}
 #'
-#' \code{\link{catboost.unify}} for \code{\link[catboost:catboost.train]{CatBoost models}}
-#'
 #' \code{\link{xgboost.unify}} for \code{\link[xgboost:xgboost]{XGBoost models}}
 #'
 #' \code{\link{randomForest.unify}} for \code{\link[randomForest:randomForest]{randomForest models}}

diff --git a/R/unify_ranger_surv.R b/R/unify_ranger_surv.R
@@ -26,8 +26,6 @@
 #'
 #' \code{\link{gbm.unify}} for \code{\link[gbm:gbm]{GBM models}}
 #'
-#' \code{\link{catboost.unify}} for \code{\link[catboost:catboost.train]{CatBoost models}}
-#'
 #' \code{\link{xgboost.unify}} for \code{\link[xgboost:xgboost]{XGBoost models}}
 #'
 #' \code{\link{randomForest.unify}} for \code{\link[randomForest:randomForest]{randomForest models}}

diff --git a/R/unify_xgboost.R b/R/unify_xgboost.R
@@ -16,8 +16,6 @@
 #'
 #' \code{\link{gbm.unify}} for \code{\link[gbm:gbm]{GBM models}}
 #'
-#' \code{\link{catboost.unify}} for  \code{\link[catboost:catboost.train]{CatBoost models}}
-#'
 #' \code{\link{ranger.unify}} for \code{\link[ranger:ranger]{ranger models}}
 #'
 #' \code{\link{randomForest.unify}} for \code{\link[randomForest:randomForest]{randomForest models}}

diff --git a/README.Rmd b/README.Rmd
@@ -21,7 +21,7 @@ set.seed(21)
 
 <!-- badges: end -->
 
-In the era of complicated classifiers conquering their market, sometimes even the authors of algorithms do not know the exact manner of building a tree ensemble model. The difficulties in models' structures are one of the reasons why most users use them simply like black-boxes. But, how can they know whether the prediction made by the model is reasonable? `treeshap` is an efficient answer for this question. Due to implementing an optimized algorithm for tree ensemble models (called TreeSHAP), it calculates the SHAP values in polynomial (instead of exponential) time. Currently, `treeshap` supports models produced with `xgboost`, `lightgbm`, `gbm`, `catboost`, `ranger`, and `randomForest` packages.
+In the era of complicated classifiers conquering their market, sometimes even the authors of algorithms do not know the exact manner of building a tree ensemble model. The difficulties in models' structures are one of the reasons why most users use them simply like black-boxes. But, how can they know whether the prediction made by the model is reasonable? `treeshap` is an efficient answer for this question. Due to implementing an optimized algorithm for tree ensemble models (called TreeSHAP), it calculates the SHAP values in polynomial (instead of exponential) time. Currently, `treeshap` supports models produced with `xgboost`, `lightgbm`, `gbm`, `ranger`, and `randomForest` packages. Support for `catboost` is available only in [`catboost` branch](https://github.com/ModelOriented/treeshap/tree/catboost) (see why [here](#catboost)).
 
 ## Installation
 
@@ -134,16 +134,15 @@ Dataset used as a reference for calculating SHAP values is stored in unified mod
 
 ```{r set_reference_dataset, eval=FALSE}
 library(treeshap)
-library(catboost)
-data <- fifa20$data[colnames(fifa20$data) != 'work_rate']
-label <- fifa20$target
-dt.pool <- catboost::catboost.load_pool(data = as.data.frame(lapply(data, as.numeric)), label = label)
-cat_model <- catboost::catboost.train(
-            dt.pool,
-            params = list(loss_function = 'RMSE', iterations = 100,
-                          logging_level = 'Silent', allow_writing_files = FALSE))
-unified_catboost <- unify(cat_model, dt.pool, data)
-unified_catboost2 <- set_reference_dataset(unified_catboost, data[c(1000:2000), ])
+library(ranger)
+data_fifa <- fifa20$data[!colnames(fifa20$data) %in%
+                             c('work_rate', 'value_eur', 'gk_diving', 'gk_handling',
+                              'gk_kicking', 'gk_reflexes', 'gk_speed', 'gk_positioning')]
+data <- na.omit(cbind(data_fifa, target = fifa20$target))
+rf <- ranger::ranger(target~., data = data, max.depth = 10, num.trees = 10)
+
+unified_ranger_model <- unify(rf, data)
+unified_ranger_model2 <- set_reference_dataset(unified_ranger_model, data[c(1000:2000), ])
 ```
 
 ## Other functionalities
@@ -158,6 +157,16 @@ Our implementation works at a speed comparable to the original Lundberg's Python
 
 The complexity of SHAP interaction values computation is $\mathcal{O}(MTLD^2)$, where $M$ is the number of explanatory variables used by the explained model, $T$ is the number of trees, $L$ is the number of leaves in a tree, and $D$ is the depth of a tree.
 
+## CatBoost
+Originally, `treeshap` also supported the CatBoost models from the `catboost` package but due to the lack of this package on CRAN or R-universe (see `catboost` issues issues [#439](https://github.com/catboost/catboost/issues/439), [#1846](https://github.com/catboost/catboost/issues/1846)), we decided to remove support from the main version of our package. 
+
+However, you can still use the `treeshap` implementation for `catboost` by installing our package from [`catboost` branch](https://github.com/ModelOriented/treeshap/tree/catboost).
+
+This branch can be installed with:
+
+``` r
+devtools::install_github('ModelOriented/treeshap@catboost')
+```
 
 ## References
 - Lundberg, S.M., Erion, G., Chen, H. et al. "From local explanations to global understanding with explainable AI for trees", Nature Machine Intelligence 2, 56–67 (2020).