SVMBench: SVM Benchmarking Data and Consensus Relations

In relations: Data Structures and Algorithms for Relations

SVM Benchmarking Data and Consensus Relations

Description

SVM_Benchmarking_Classification and SVM_Benchmarking_Regression represent the results of a benchmark study (Meyer, Leisch and Hornik, 2003) comparing Support Vector Machines to other predictive methods on real and artificial data sets involving classification and regression methods, respectively. In addition, SVM_Benchmarking_Classification_Consensus and SVM_Benchmarking_Regression_Consensus provide consensus rankings derived from these data.

Usage

data("SVM_Benchmarking_Classification")
data("SVM_Benchmarking_Regression")
data("SVM_Benchmarking_Classification_Consensus")
data("SVM_Benchmarking_Regression_Consensus")

Format

SVM_Benchmarking_Classification (SVM_Benchmarking_Regression) is an ensemble of 21 (12) relations representing pairwise comparisons of 17 classification (10 regression) methods on 21 (12) data sets. Each relation of the ensemble summarizes the results for a particular data set. The relations are reflexive endorelations on the set of methods employed, with a pair (a, b) of distinct methods contained in a relation iff both delivered results on the corresponding data set and a did not perform significantly better than b at the 5% level. Since some methods failed on some data sets, the relations are not guaranteed to be complete or transitive. See Meyer et al. (2003) for details on the experimental design of the benchmark study, and Hornik and Meyer (2007) for the pairwise comparisons.

The corresponding consensus objects are lists of ensembles of consensus relations fitted to the benchmark results. For each of the following three endorelation families: SD/L (“linear orders”), SD/O (“partial orders”), and SD/W (“weak orders”), all possible consensus relations have been computed (see relation_consensus). For both classification and regression, the three relation ensembles obtained are provided as a named list of length 3. See Hornik & Meyer (2007) for details on the meta-analysis.

Source

D. Meyer, F. Leisch, and K. Hornik (2003), The support vector machine under test. Neurocomputing, 55, 169–186. doi: 10.1016/S0925-2312(03)00431-4.

K. Hornik and D. Meyer (2007), Deriving consensus rankings from benchmarking experiments. In R. Decker and H.-J. Lenz, Advances in Data Analysis. Studies in Classification, Data Analysis, and Knowledge Organization. Springer-Verlag: Heidelberg, 163–170.

Examples

data("SVM_Benchmarking_Classification")

## 21 data sets
names(SVM_Benchmarking_Classification)

## 17 methods
relation_domain(SVM_Benchmarking_Classification)

## select weak orders
weak_orders <-
    Filter(relation_is_weak_order, SVM_Benchmarking_Classification)

## only the artifical data sets yield weak orders
names(weak_orders)

## visualize them using Hasse diagrams
if(require("Rgraphviz")) plot(weak_orders)

## Same for regression:
data("SVM_Benchmarking_Regression")

## 12 data sets
names(SVM_Benchmarking_Regression)

## 10 methods
relation_domain(SVM_Benchmarking_Regression)

## select weak orders
weak_orders <-
    Filter(relation_is_weak_order, SVM_Benchmarking_Regression)

## only two of the artifical data sets yield weak orders
names(weak_orders)

## visualize them using Hasse diagrams
if(require("Rgraphviz")) plot(weak_orders)

## Consensus solutions:

data("SVM_Benchmarking_Classification_Consensus")
data("SVM_Benchmarking_Regression_Consensus")

## The solutions for the three families are not unique
print(SVM_Benchmarking_Classification_Consensus)
print(SVM_Benchmarking_Regression_Consensus)

## visualize the consensus weak orders
classW <- SVM_Benchmarking_Classification_Consensus$W
regrW <- SVM_Benchmarking_Regression_Consensus$W
if(require("Rgraphviz")) {
    plot(classW)
    plot(regrW)
}

## in tabular style:
ranking <- function(x) rev(names(sort(relation_class_ids(x))))
sapply(classW, ranking)
sapply(regrW, ranking)

## (prettier and more informative:)
relation_classes(classW[[1L]])

relations documentation built on March 7, 2023, 8:01 p.m.