xgrove (version 0.1-15)

xgrove: Explanation groves

Description

Compute surrogate groves to explain predictive machine learning model and analyze complexity vs. explanatory power.

Usage

xgrove(
  model,
  data,
  ntrees = c(4, 8, 16, 32, 64, 128),
  pfun = NULL,
  remove.target = T,
  shrink = 1,
  b.frac = 1,
  seed = 42,
  ...
)

Value

List of the results:

explanation

Matrix containing tree sizes, rules, explainability \({\Upsilon}\) and the correlation between the predictions of the explanation and the true model.

rules

Summary of the explanation grove: Rules with identical splits are aggegated. For numeric variables any splits are merged if they lead to identical parititions of the training data.

groves

Rules of the explanation grove.

model

gbm model.

Arguments

model

A model with corresponding predict function that returns numeric values.

data

Training data.

ntrees

Sequence of integers: number of boosting trees for rule extraction.

pfun

Optional predict function function(model, data) returning a real number. Default is the predict() method of the model.

remove.target

Logical. If TRUE the name of the target variable is identified from terms(model) and automatically removed if this variable is still in data.

shrink

Sets the shrinkage argument for the internal call of gbm. As the model usually has a deterministic response the default is 1 different to the default of gbm applied train a model based on data.

b.frac

Sets the bag.fraction argument for the internal call of gbm. As the model usually has a deterministic response the default is 1 different to the default of gbm applied train a model based on data.

seed

Seed for the random number generator to ensure reproducible results (e.g. for the default bag.fraction < 1 in boosting).

...

Further arguments to be passed to gbm or the predict() method of the model.

Author

Details

A surrogate grove is trained via gradient boosting using gbm on data with the predictions of using of the model as target variable. Note that data must not contain the original target variable! The boosting model is trained using stumps of depth 1. The resulting interpretation is extracted from pretty.gbm.tree. The column upper_bound_left of the rules and the groves value of the output object contains the split point for numeric variables denoting the uppoer bound of the left branch. Correspondingly, the levels_left column contains the levels of factor variables assigned to the left branch. The rule weights of the branches are given in the rightmost columns. The prediction of the grove is obtained as the sum of the assigned weights over all rows. Note that the training data must not contain the target variable. It can be either removed manually or will be removed automatically from data if the argument remove.target == TRUE.

References

• Szepannek, G. and von Holt, B.H. (2023): Can’t see the forest for the trees -- analyzing groves to explain random forests, Behaviormetrika, DOI: 10.1007/s41237-023-00205-2.

• Szepannek, G. and Luebke, K.(2023): How much do we see? On the explainability of partial dependence plots for credit risk scoring, Argumenta Oeconomica 50, DOI: 10.15611/aoe.2023.1.07.

Examples

library(randomForest)
library(pdp)
data(boston)
set.seed(42)
rf <- randomForest(cmedv ~ ., data = boston)
data <- boston[,-3] # remove target variable
ntrees <- c(4,8,16,32,64,128)
xg <- xgrove(rf, data, ntrees)
xg
plot(xg)

# Example of a classification problem using the iris data.
# A predict function has to be defined, here for the posterior probabilities of the class Virginica.  
data(iris)
set.seed(42)
rf    <- randomForest(Species ~ ., data = iris)
data  <- iris[,-5] # remove target variable

pf <- function(model, data){
  predict(model, data, type = "prob")[,3]
  }
  
xgrove(rf, data, pfun = pf)