Plot Pair-wise Overlap of Ranked Features
rankingPlot.Rd
Pair-wise overlaps can be done for two types of analyses. Firstly, each cross-validation iteration can be considered within a single classification. This explores the feature ranking stability. Secondly, the overlap may be considered between different classification results. This approach compares the feature ranking commonality between different results. Two types of commonality are possible to analyse. One summary is the average pair-wise overlap between all possible pairs of results. The second kind of summary is the pair-wise overlap of each level of the comparison factor that is not the reference level against the reference level. The overlaps are converted to percentages and plotted as lineplots.
Usage
# S4 method for ClassifyResult
rankingPlot(results, ...)
# S4 method for list
rankingPlot(
results,
topRanked = seq(10, 100, 10),
comparison = "within",
referenceLevel = NULL,
characteristicsList = list(),
orderingList = list(),
sizesList = list(lineWidth = 1, pointSize = 2, legendLinesPointsSize = 1, fonts = c(24,
16, 12, 12, 12, 16)),
lineColours = NULL,
xLabelPositions = seq(10, 100, 10),
yMax = 100,
title = if (comparison[1] == "within") "Feature Ranking Stability" else
"Feature Ranking Commonality",
yLabel = if (is.null(referenceLevel)) "Average Common Features (%)" else
paste("Average Common Features with", referenceLevel, "(%)"),
margin = grid::unit(c(1, 1, 1, 1), "lines"),
showLegend = TRUE,
parallelParams = bpparam()
)
Arguments
- results
A list of
ClassifyResult
objects.- ...
Not used by end user.
- topRanked
A sequence of thresholds of number of the best features to use for overlapping.
- comparison
Default:
"within"
. The aspect of the experimental design to compare. Can be any characteristic that all results share or special value "within" to compared between all pairwise iterations of cross-validation.- referenceLevel
The level of the comparison factor to use as the reference to compare each non-reference level to. If
NULL
, then each level has the average pairwise overlap calculated to all other levels.- characteristicsList
A named list of characteristics. The name must be one of
"lineColour"
,"pointType"
,"row"
or"column"
. The value of each element must be a characteristic name, as stored in the"characteristic"
column of the results' characteristics table.- orderingList
An optional named list. Any of the variables specified to
characteristicsList
can be the name of an element of this list and the value of the element is the order in which the factor should be presented in.- sizesList
Default:
lineWidth = 1, pointSize = 2, legendLinesPointsSize = 1, fonts = c(24, 16, 12, 12, 12, 16)
. A list which must contain elements namedlineWidth
,pointSize
,legendLinesPointsSize
andfonts
. The first three specify the size of lines and points in the graph, as well as in the plot legend.fonts
is a vector of length 6. The first element is the size of the title text. The second element is the size of the axes titles. The third element is the size of the axes values. The fourth element is the size of the legends' titles. The fifth element is the font size of the legend labels. The sixth element is the font size of the titles of grouped plots, if any are produced. Each list element must numeric.- lineColours
A vector of colours for different levels of the line colouring parameter, if one is specified by
characteristicsList[["lineColour"]]
. If none are specified but,characteristicsList[["lineColour"]]
is, an automatically-generated palette will be used.- xLabelPositions
Locations where to put labels on the x-axis.
- yMax
The maximum value of the percentage to plot.
- title
An overall title for the plot.
- yLabel
Label to be used for the y-axis of overlap percentages.
- margin
The margin to have around the plot.
- showLegend
If
TRUE
, a legend is plotted next to the plot. If FALSE, it is hidden.- parallelParams
An object of class
MulticoreParam
orSnowParam
.
Details
If comparison
is "within"
, then the feature selection overlaps
are compared within a particular analysis. The result will inform how stable
the selections are between different iterations of cross-validation for a
particular analysis. Otherwise, the comparison is between different
cross-validation runs, and this gives an indication about how common are the
features being selected by different classifications.
Calculating all pair-wise set overlaps for a large cross-validation result
can be time-consuming. This stage can be done on multiple CPUs by providing
the relevant options to parallelParams
.
Examples
predicted <- DataFrame(sample = sample(10, 100, replace = TRUE),
permutation = rep(1:2, each = 50),
class = rep(c("Healthy", "Cancer"), each = 50))
actual <- factor(rep(c("Healthy", "Cancer"), each = 5))
allFeatures <- sapply(1:100, function(index) paste(sample(LETTERS, 3), collapse = ''))
rankList <- list(allFeatures[1:100], allFeatures[c(15:6, 1:5, 16:100)],
allFeatures[c(1:9, 11, 10, 12:100)], allFeatures[c(1:50, 61:100, 60:51)])
result1 <- ClassifyResult(DataFrame(characteristic = c("Data Set", "Selection Name", "Classifier Name", "Cross-validation"),
value = c("Melanoma", "t-test", "Diagonal LDA", "2 Permutations, 2 Folds")),
LETTERS[1:10], allFeatures, rankList,
list(rankList[[1]][1:15], rankList[[2]][1:15],
rankList[[3]][1:10], rankList[[4]][1:10]),
list(function(oracle){}), NULL,
predicted, actual)
predicted[, "class"] <- sample(predicted[, "class"])
rankList <- list(allFeatures[1:100], allFeatures[c(sample(20), 21:100)],
allFeatures[c(1:9, 11, 10, 12:100)], allFeatures[c(1:50, 60:51, 61:100)])
result2 <- ClassifyResult(DataFrame(characteristic = c("Data Set", "Selection Name", "Classifier Name",
"Cross-validations"),
value = c("Melanoma", "t-test", "Random Forest", "2 Permutations, 2 Folds")),
LETTERS[1:10], allFeatures, rankList,
list(rankList[[1]][1:15], rankList[[2]][1:15],
rankList[[3]][1:10], rankList[[4]][1:10]),
list(function(oracle){}), NULL,
predicted, actual)
rankingPlot(list(result1, result2), characteristicsList = list(pointType = "Classifier Name"))