imptree-package {imptree}

imptree: Classification Trees with Imprecise Probabilities

Description

The imptree package implements the creation of imprecise classification trees based on algorithm developed by Abellan and Moral. The credal sets of the classification variable within each node are estimated by either the imprecise Dirichlet model (IDM) or the nonparametric predictive inference (NPI). As split possible split criteria serve the 'information gain', based on the maximal entropy distribution, and the adaptable entropy-range based criterion propsed by Fink and Crossman. It also implements different correction terms for the entropy.

The performance of the tree can be evaluated with respect to the common criteria in the context of imprecise classification trees.

It also provides the functionality for estimating credal sets via IDM or NPI and obtain their minimal/maximal entropy (distribution) to be used outside the tree growing process.

References

Abellán, J. and Moral, S. (2005), Upper entropy of credal sets. Applications to credal classification, International Journal of Approximate Reasoning 39, pp. 235–255.

Baker, R. M. (2010), Multinomial Nonparametric Predictive Inference: Selection, Classification and Subcategory Data, PhD thesis. Durham University, GB.

Strobl, C. (2005), Variable Selection in Classification Trees Based on Imprecise Probabilities, ISIPTA '05: Proceedings of the Fourth International Symposium on Imprecise Probabilities and Their Applications, 339–348.

Fink, P. and Crossman, R.J. (2013), Entropy based classification trees, ISIPTA '13: Proceedings of the Eighth International Symposium on Imprecise Probability: Theories and Applications, pp. 139–147.

See Also

imptree for tree creation, probInterval for the credal set and entropy estimation functionality

Examples

data("carEvaluation")

## create a tree with IDM (s=1) to full size
## carEvaluation, leaving the first 10 observations out
ip <- imptree(acceptance~., data = carEvaluation[-(1:10),], 
  method="IDM", method.param = list(splitmetric = "globalmax", s = 1), 
  control = list(depth = NULL, minbucket = 1))

## summarize the tree and show performance on training data
summary(ip)

## predict the first 10 observations
## Note: The result of the prediction is return invisibly
pp <- predict(ip, dominance = "max", data = carEvaluation[(1:10),])
## print the general evaluation statistics
print(pp)
## display the predicted class labels
pp$classes