manual_docs/BostonCIU.md

In KaryFramling/ciu: Contextual Importance and Utility

CIU with Boston Housing Dataset

The Boston Housing Dataset is a classical benchmark set used in Data Analysis and Machine learning, available in the MASS library.

A gradient boosting model will be used here for learning the mapping from the 13 input variables to the Median value of owner-occupied homes in $1000's. This is a regression task (not a classification task), which means that the best explanation approach might be different. However, for CIU it doesn't really matter whether it is a classification or a regression task.

gbm is not (yet) supported directly by ciu, so we will provide the predictfunction to use as an argument to ciu.new. This is a generic way for providing support for "any" regression/classification model.

require(MASS) # Just in case Boston is not already available
require(gbm)

# We don't care about train/test set in this case because it's not about evaluating training performance.
Boston.boost=gbm(medv ~ . ,data = Boston, distribution = "gaussian", n.trees=10000,
                 shrinkage = 0.01, interaction.depth = 4)

Gradient Boosting (gbm) is not yet supported by ciu implementation, so we need to tell ciu explicitly what it looks like.

predict.function <- function(model, inputs) { predict(model,inputs,n.trees=10000) }

We initialize and create CIU object and get CI and CU for row 370, which has a high median value (50).

n.in <- ncol(Boston) - 1
ciu <- ciu.new(Boston.boost, medv~., Boston, predict.function=predict.function)
inst.ind <- 370
ciu$explain(Boston[inst.ind,1:n.in], ind.inputs.to.explain=c(1))
# Gives CI and CU for the first input variable (crim) and looks something like this: 
#         CI        CU     cmin     cmax   outval
#1 0.1630439 0.8111716 44.55798 51.89496 50.50953

Then we display bar, ggplot.col and pie charts with the following. Example of ggplot version is shown below.

ciu$ggplot.col.ciu(Boston[inst.ind,1:n.in])
ciu$barplot.ciu(Boston[inst.ind,1:n.in])
ciu$pie.ciu(Boston[inst.ind,1:n.in])

Alternative way for supporting "any" model

The CIU.BlackBox class is supported automatically by ciu. It essentially means that the model class needs to inherit CIU.BlackBox and implement an eval(inputs)method.

The following code shows how this can be done for Gradient Boosting. The same principle is applicable for any model. This approach is more flexible than providing a predict.function argument to ciu.new() because it is possible to associate information with the object. In this example, the n.treesparameter is stored as an object instance value but it is possible to implement any required processing with this approach.

# CIU with Gradient Boosting, CIU.BlackBox wrapper
gbm.bb.new <- function(gbm, n.trees=1) {
  o.gbm <- gbm
  o.n.trees <- n.trees
  pub <- list(eval = function(inputs) { predict(o.gbm,inputs,n.trees=o.n.trees) })
  class(pub) <- c("CIU.BlackBox",class(pub))
  return(pub)
}
gbm.bb <- gbm.bb.new(Boston.boost, 10000)
ciu.gbm.bb <- ciu.new(gbm.bb, medv~., Boston)
inst.ind <- 370
ciu.gbm.bb$barplot.ciu(Boston[inst.ind,1:n.in], sort="CI")

Example barplots by CIU

The following figure shows the barplot of three instances: #406 (medv=5, lowest), #6 (medv=28.7, average), #370 (medv=50, highest).

The code for producing this image is the following:

par(mfrow=c(1,3))
ciu.gbm.bb$barplot.ciu(Boston[406,1:n.in])
ciu.gbm.bb$barplot.ciu(Boston[6,1:n.in])
ciu.gbm.bb$barplot.ciu(Boston[370,1:n.in])
par(mfrow=c(1,1))

Same as pie charts (using pie.CI.CU instead of barplot.CI.CU):

Author

Kary Främling

KaryFramling/ciu documentation built on July 16, 2025, 8:18 p.m.