pairedComparisons: Statistical hypothesis testing on the observed paired...
In performanceEstimation: An Infra-Structure for Performance Estimation of Predictive Models

Description Usage Arguments Details Value Author(s) References See Also Examples

This function analyses the statistical significance of the paired comparisons between the estimated performance scores of a set of workflows. When you run the performanceEstimation() function to compare a set of workflows over a set of problems you obtain estimates of their performances across these problems. This function implements several statistical tests that can be used to test several hypothesis concerning the observed differences in performance between the workflows on the tasks.

1
2
3

pairedComparisons(obj,baseline,
                  maxs=rep(FALSE,length(metricNames(obj))),
                  p.value=0.05)

`obj`	An object of class `ComparisonResults` that contains the results of a performance estimation experiment.
`baseline`	Several tests involve the hypothesis that a certain workflow is significantly different from a set of other alternatives. This argument allows you to specify the name of this baseline workflow. If you omit this name the function will default to the name of the workflow that has the lower average rank position across all tasks, for each estimation metric.
`maxs`	A vector of booleans with as many elements are there are metrics estimated in the experiment. A `TRUE` value means the respective metric is to be maximized, while a `FALSE` means minimization. Defaults to all `FALSE` values, i.e. all metrics are to be minimized.
`p.value`	A p value to be used in the calculations that involve using values from statistical tables (defaults to 0.05).

The performanceEstimation function allows you to obtain estimates of the expected value of a series of performance metrics for a set of alternative workflows and a set of predictive tasks. After running this type of experiments we frequently want to check if there is any statistical significance between the estimated performance of the different workflows. The current function allows you to carry out this type of checks.

The function will only run on experiments containing more than one workflow as paired comparisons do not make sense with one single alternative workflow. Having more than one workflow we can distinguish two situations: i) comparing the performance of two workflows; or ii) comparisons among multiple workflows. The recommendations for checking the statistical significance of the difference between the performance of the alternative workflows varies within these two setups (see Demsar (2006) for recommendations).

The current function implements several statistical tests that can be used for different hypothesis tests. Namely, it obtains the results of paired t tests and paired Wilcoxon Signed Rank tests for situations where you are comparing the performance of two workflows, with the latter being recommended given the typical overlap among the training sets that does not ensure independence among the scores of the different iterations. For the setup of multiple workflows on multiple tasks the function also calculates the Friedman test and the post-hoc Nemenyi and Bonferroni-Dunn tests, according to the procedures described in Demsar (2006). The combination Friedman test followed by the post-hoc Nemenyi test is recommended when you want to carry out paired comparisons between all alternative workflows on the set of tasks to check for which differences are significant. The combination Friedman test followed by the post-hoc Bonferroni-Dunn test is recommended when you want to compare a set of alternative workflows against a baseline workflow. For both of these two paths we provide an implementation of the diagrams (CD diagrams) described in Demsar (2006) through the functions CDdiagram.BD and CDdiagram.Nemenyi.

The performanceEstimation function ensures that all compared workflows are run on exactly the same train+test partitions on all repetitions and for all predictive tasks. In this context, we can use pairwise statistical significance tests.

The result of this function is the information from performing all these statistical tests. This information is returned as a list with as many components as there are estimated metrics. For each metric a list with several components describing the results of these tests is provided.

Luis Torgo ltorgo@dcc.fc.up.pt

Demsar, J. (2006) Statistical Comparisons of Classifiers over Multiple Data Sets. Journal of Machine Learning Research, 7, 1-30.

Torgo, L. (2014) An Infra-Structure for Performance Estimation and Experimental Comparison of Predictive Models in R. arXiv:1412.0436 [cs.MS] http://arxiv.org/abs/1412.0436

CDdiagram.Nemenyi, CDdiagram.BD, signifDiffs, performanceEstimation, topPerformers, topPerformer, rankWorkflows, metricsSummary, ComparisonResults

## Not run: 
## Estimating MSE for 3 variants of both
## regression trees and SVMs, on  two data sets, using one repetition
## of 10-fold CV
library(e1071)
data(iris)
data(Satellite,package="mlbench")
data(LetterRecognition,package="mlbench")


## running the estimation experiment
res <- performanceEstimation(
           c(PredTask(Species ~ .,iris),
             PredTask(classes ~ .,Satellite,"sat"),
             PredTask(lettr ~ .,LetterRecognition,"letter")),
           workflowVariants(learner="svm",
                 learner.pars=list(cost=1:4,gamma=c(0.1,0.01))),
           EstimationTask(metrics=c("err","acc"),method=CV()))


## checking the top performers
topPerformers(res)

## now let us assume that we will choose "svm.v2" as our baseline
## carry out the paired comparisons
pres <- pairedComparisons(res,"svm.v2")

## obtaining a CD diagram comparing all others against svm.v2 in terms
## of error rate
CDdiagram.BD(pres,metric="err")


## End(Not run)