ciu.textual: Give textual CIU explanation

Description

Provide textual CIU explanations as those used in Kary Främling's PhD thesis.

Usage

ciu.textual(
  ciu,
  instance = NULL,
  ind.inputs = NULL,
  ind.output = 1,
  in.min.max.limits = NULL,
  n.samples = 100,
  neutral.CU = 0.5,
  show.input.values = TRUE,
  concepts.to.explain = NULL,
  target.concept = NULL,
  target.ciu = NULL,
  ciu.meta = NULL,
  sort = "CI",
  n.features = NULL,
  use.text.effects = FALSE,
  CI.voc = data.frame(limits = c(0.2, 0.4, 0.6, 0.8, 1), texts = c("not important",
    "slightly important", "important", "very important", "extremely important")),
  CU.voc = data.frame(limits = c(0.2, 0.4, 0.6, 0.8, 1), texts = c("very bad", "bad",
    "average", "good", "very good"))
)

Value

Text string with explanation.

Arguments

ciu: ciu object as created with ciu function (not to be confused with CIU object as created by ciu.new).
instance: Input values for the instance to explain. Should be a data.frame even though a vector or matrix might work too if input names and other needed metadata can be deduced from the dataset or other parameters given to ciu.new.
ind.inputs: Indices of input features to explain (the set i in CIU formulae)
ind.output: Index of output to be explained.
in.min.max.limits: data.frame or matrix with one row per output and two columns, where the first column indicates the minimal value and the second column the maximal value for that output. ONLY NEEDED HERE IF not given as parameter to ciu.new or if the limits are different for this specific instance than the default ones.
n.samples: How many instances to generate for estimating CI and CU. For inputs of type factor, all possible combinations of input values are generated, so this parameter only influences how many instances are (at least) generated for continuous-valued inputs.
neutral.CU: Indicates when the Contextual Utility is considered to be "negative". The default value of 0.5 seems quite logical for most cases.
show.input.values: Include input values after input labels or not. Default is TRUE.
concepts.to.explain: List of input feature concepts to explain, as defined by vocabulary provided as argument to ciu.new. If ind.inputs=NULL, then use concepts.to.explain instead. If both are NULL, then use all inputs.
target.concept: If provided, then calculate CIU of inputs ind.inputs.to.explain relative to the given concept rather than relative to the actual output(s). ind.inputs.to.explain should normally be a subset (or all) of the inputs that target.concept consists of, even though that not required by the CIU calculation. If a target.ciu is provided, then the target.concept doesn't have to be included in the vocabulary gives as parameter to ciu.new (at least for the moment).
target.ciu: ciu.result object previously calculated for target.concept. If a target.concept is provided but target.ciu=NULL, then target.ciu is estimated by a call to ciu.explain with the n.samples value given as a parameter to this call. It may be useful to provide target.ciu if it should be estimated using some other (typically greater) value for n.samples than the default one, or if it has already been calculated for some reason.
ciu.meta: If given, then use existing ciu.meta.result rather than calling ciu.meta.explain.
sort: NULL, "CI" or "CU".
n.features: Maximal number of features to include in explanation.
use.text.effects: Add bold/italics/colors effects?
CI.voc: Limits and texts to use for CI values.
CU.voc: Limits and texts to use for CU values.

Examples

Run this code

# Explaining the classification of an Iris instance with lda model.
# We use a versicolor (instance 100).
library(MASS)
test.ind <- 100
iris_test <- iris[test.ind, 1:4]
iris_train <- iris[-test.ind, 1:4]
iris_lab <- iris[[5]][-test.ind]
model <- lda(iris_train, iris_lab)

# Create CIU object
ciu <- ciu.new(model, Species~., iris)

# Give textual explanation. Use 'cat' for getting newlines to work.
cat(ciu.textual(ciu, iris_test, ind.output = 2))
cat(ciu.textual(ciu, iris_test, ind.output = 2, n.features = 2))

# Boston housing, GBM model.
library(caret)
kfoldcv <- trainControl(method="cv", number=10)
gbm <- train(medv ~ ., Boston, method="gbm", trControl=kfoldcv)
boston.inst <- Boston[370,1:13]
ciu <- ciu.new(gbm, medv~., Boston)
cat(ciu.textual(ciu, boston.inst,use.text.effects = TRUE))
# Customized limits for CI.
cat(ciu.textual(ciu, boston.inst,use.text.effects = TRUE,
  CI.voc = data.frame(limits=c(0.05,0.1,0.3,0.5,1.0),
texts=c("not important","little important", "important","very important",
  "extremely important"))))

# Intermediate concepts
social<-c(1,11,13); usage_type<-c(2,3); chas<-c(4); air_quality<-c(5)
housing<-c(6,7); transport<-c(8,9); blacks<-c(12); tax<-c(10)
Boston.voc <- list("SOCIAL"=social, "LAND USAGE"=usage_type, "Charles R. dummy"=chas,
"Air quality (Nox)"=air_quality, "HOUSING"=housing, "TRANSPORT"=transport,
"Prop. of black people"=blacks, "Tax"=tax)
ciu <- ciu.new(gbm, medv~., Boston, vocabulary = Boston.voc)

# We use `meta.explain` here to avoid differences due to sampling.
meta.top <- ciu$meta.explain(boston.inst, concepts.to.explain=names(Boston.voc))
cat(ciu.textual(ciu, boston.inst, use.text.effects = TRUE, ciu.meta = meta.top))

# Explain intermediate concept utility, using input features (could also
# be using other intermediate concepts).
cat(ciu.textual(ciu, boston.inst, use.text.effects = TRUE, ind.inputs = Boston.voc$SOCIAL,
  target.concept = "SOCIAL", target.ciu = meta.top$ciuvals[["SOCIAL"]]))
cat(ciu.textual(ciu, boston.inst, use.text.effects = TRUE, ind.inputs = Boston.voc$HOUSING,
  target.concept = "HOUSING", target.ciu = meta.top$ciuvals[["HOUSING"]]))

Run the code above in your browser using DataLab