mlr (version 2.12.1)

generateCritDifferencesData: Generate data for critical-differences plot.

Description

Generates data that can be used to plot a critical differences plot. Computes the critical differences according to either the `"Bonferroni-Dunn"` test or the `"Nemenyi"` test. `"Bonferroni-Dunn"` usually yields higher power as it does not compare all algorithms to each other, but all algorithms to a `baseline` instead. Learners are drawn on the y-axis according to their average rank. For `test = "nemenyi"` a bar is drawn, connecting all groups of not significantly different learners. For `test = "bd"` an interval is drawn arround the algorithm selected as baseline. All learners within this interval are not signifcantly different from the baseline. Calculation: $$ CD = q_{\alpha} \sqrt{(\frac{k(k+1)}{6N})}$$ Where \(q_\alpha\) is based on the studentized range statistic. See references for details.

Usage

generateCritDifferencesData(bmr, measure = NULL, p.value = 0.05,
  baseline = NULL, test = "bd")

Arguments

bmr

(BenchmarkResult) Benchmark result.

measure

(Measure) Performance measure. Default is the first measure used in the benchmark experiment.

p.value

([numeric`(1)] P-value for the critical difference. Default: 0.05

baseline

(`character(1)`): ([learner.id]) Select a `learner.id` as baseline for the `test = "bd"` ("Bonferroni-Dunn") critical differences diagram.The critical difference Interval will then be positioned arround this learner. Defaults to best performing algorithm. For `test = "nemenyi"`, no baseline is needed as it performs `all pairwise comparisons.`

test

(`character(1)`) Test for which the critical differences are computed. “bd” for the Bonferroni-Dunn Test, which is comparing all classifiers to a `baseline`, thus performing a comparison of one classifier to all others. Algorithms not connected by a single line are statistically different. then the baseline. “nemenyi” for the [PMCMR::posthoc.friedman.nemenyi.test] which is comparing all classifiers to each other. The null hypothesis that there is a difference between the classifiers can not be rejected for all classifiers that have a single grey bar connecting them.

Value

([critDifferencesData]). List containing:

data

([data.frame]) containing the info for the descriptive part of the plot

friedman.nemenyi.test

([list]) of class `pairwise.htest` contains the calculated posthoc.friedman.nemenyi.test

cd.info

([list]) containing info on the critical difference and its positioning

baseline

`baseline` chosen for plotting

p.value

p.value used for the posthoc.friedman.nemenyi.test and for computation of the critical difference

See Also

Other generate_plot_data: generateCalibrationData, generateFeatureImportanceData, generateFilterValuesData, generateLearningCurveData, generatePartialDependenceData, generateThreshVsPerfData, getFilterValues, plotFilterValues

Other benchmark: BenchmarkResult, batchmark, benchmark, convertBMRToRankMatrix, friedmanPostHocTestBMR, friedmanTestBMR, getBMRAggrPerformances, getBMRFeatSelResults, getBMRFilteredFeatures, getBMRLearnerIds, getBMRLearnerShortNames, getBMRLearners, getBMRMeasureIds, getBMRMeasures, getBMRModels, getBMRPerformances, getBMRPredictions, getBMRTaskDescs, getBMRTaskIds, getBMRTuneResults, plotBMRBoxplots, plotBMRRanksAsBarChart, plotBMRSummary, plotCritDifferences, reduceBatchmarkResults