Lift Chart

Draws the cumulative lift curve, the cumulative gain curve, or the per-group lift bars of a binary classifier, together with the baseline of a random model. The chart answers the operational question directly: how much better is acting on the top-scored share of the cases than acting on a random share of the same size.

Usage

plotLift(
  x,
  main = NULL,
  xlab = NULL,
  ylab = NULL,
  ylim = NULL,
  type = c("cumulative", "gain", "decile"),
  col = .useTheme,
  lwd = 2,
  grid = .useTheme,
  box = .useTheme,
  baseline = TRUE,
  perfect = FALSE,
  legend = TRUE,
  stamp = .useTheme,
  ...
)

Arguments

x

an object of class "Lift", as returned by alloy::lift().

main

main title. NULL (default) draws none, unless the theme sets autoTitle = TRUE (see theme); the title is then derived from the expression passed as x. "", NA or FALSE always suppress it and compact the top margin; any other string is used as given.

xlab

label for the x-axis. NULL (default) derives a label from type.

ylab

label for the y-axis. NULL (default) derives a label from type.

ylim

numeric vector of length 2; y-axis limits. NULL (default) spans the curve together with the baseline.

type

the curve to draw. One of "cumulative" (cumulative lift over depth, the default), "gain" (share of all positives captured, over depth), or "decile" (per-group lift as bars).

col

color of the curve or the bars. .useTheme (default) resolves to getTheme()$twin[1] - a single accent color, consistent with plotECDF().

lwd

line width of the curve. Has no effect for type = "decile".

grid

controls drawing of the background grid. Can be:

• .useTheme (default): follow the active theme (getTheme()$grid)

• TRUE: draw grid with theme settings

• FALSE, NULL, or NA: suppress grid

• a named list: arguments passed to graphics::grid(), overriding the theme defaults for this call only

box

controls drawing of the plot box. .useTheme (default) resolves to getTheme()$box. TRUE/FALSE/NA, or a named list, as for grid.

baseline

controls the reference line of a random model - a horizontal line at 1 for type = "cumulative" and "decile", the diagonal for type = "gain". Can be:

• TRUE (default): draw with default settings

• FALSE, NULL, or NA: suppress

• a named list: arguments passed to graphics::lines() (or graphics::abline() for type = "decile"), e.g. list(col = "black", lty = "dotted")

perfect

controls the curve of a perfect ranking - the theoretical maximum attainable at each depth, given the base rate. FALSE by default, since it compresses the interesting part of the y-axis; TRUE or a named list to draw it, as for baseline. Has no effect for type = "decile".

legend

controls drawing of the legend. Can be:

• TRUE (default): draw with default settings

• FALSE, NULL, or NA: suppress

• a named list: arguments passed to graphics::legend(), e.g. list(x = "bottomleft")

stamp

controls the corner stamp. .useTheme (default) resolves to getTheme()$stamp. TRUE/FALSE/ NULL, a string, or a named list of arguments for stamp().

...

further graphical parameters passed to par() via the internal framework.

Value

Invisibly returns x.

Details

Reading the cumulative curve at depth 0.2 gives the factor by which contacting the top-scored fifth of the cases beats contacting a random fifth. The gain variant answers the complementary question - what share of all positives that fifth captures.

The curve necessarily converges to 1 (cumulative lift) or to the diagonal endpoint (gain) at depth 1, where all cases are selected and the model no longer discriminates. Only the left part of the curve carries decision value: a model that separates well over the first two deciles and poorly afterwards is preferable for a small campaign to one with the reverse profile and identical AUC.

Optional plot components (grid, box, baseline, perfect, legend) follow bedrock::callIf() semantics:

• TRUE: draw with defaults

• FALSE, NULL, or NA: suppress component

• named list: customize component arguments

col, grid, box, and stamp default to .useTheme, deferring to the package's active theme (see theme) rather than a hardcoded value.

The number of groups is a property of the lift table, not of the plot - set it via the nBins argument of alloy::lift().

See also

alloy::lift(), alloy::roc(), plotECDF(), bedrock::callIf(), theme

Other plot.special: plotBinaryTree(), plotCirc(), plotMiss(), plotPolar(), plotPropCI(), plotSankey(), plotTernary(), plotTimeSeries(), plotTreemap(), plotWeb()

Examples

# A lift table is normally computed by alloy::lift(). To keep the example
# self-contained, the columns used by the plot are built by hand here:
# cases ranked by decreasing score and cut into equally sized groups.
liftTable <- function(score, y, nBins = 10) {
  y    <- y[order(score, decreasing = TRUE)]
  bin  <- cut(seq_along(y), nBins, labels = FALSE)
  nObs <- as.vector(table(bin))
  nPos <- as.vector(tapply(y, bin, sum))
  structure(
    data.frame(bin     = seq_len(nBins),
               lift    = (nPos / nObs) / mean(y),
               depth   = cumsum(nObs) / length(y),
               cumLift = (cumsum(nPos) / cumsum(nObs)) / mean(y),
               gain    = cumsum(nPos) / sum(y)),
    class    = c("Lift", "data.frame"),
    baseRate = mean(y))
}

y   <- as.integer(iris$Species == "virginica")
fit <- glm(y ~ Sepal.Length + Sepal.Width, data = iris, family = binomial)
lft <- liftTable(fitted(fit), y)

plotLift(lft)

plotLift(lft, type = "gain")


# Per-group bars, coarser grouping set at computation time
plotLift(liftTable(fitted(fit), y, nBins = 5), type = "decile")


# Add the perfect-ranking reference, suppress the legend
plotLift(lft, perfect = TRUE, legend = FALSE)


# No title, compact top margin
plotLift(lft, main = "")