plot_gbm {plotmo}

Plot a gbm model

Description

Plot a gbm model showing the training and other error curves.

Usage

plot_gbm(object=stop("no 'object' argument"),
    smooth = c(0, 0, 0, 1),
    col = c(1, 2, 3, 4), ylim = "auto",
    legend.x = NULL, legend.y = NULL, legend.cex = .8,
    grid.col = NA,
    n.trees = NA, col.n.trees ="darkgray",
    ...)

Arguments

object

The gbm model.

smooth

Four-element vector specifying if smoothing should be applied to the train, test, CV, and OOB curves respectively. When smoothing is specified, a smoothed curve is plotted and the minimum is calculated from the smoothed curve.
The default is c(0, 0, 0, 1) meaning apply smoothing only to the OOB curve (same as gbm.perf).
Note that smooth=1 (which gets recyled to c(1,1,1,1)) will smooth all the curves.

col

Four-element vector specifying the colors for the train, test, CV, and OOB curves respectively.
The default is c(1, 2, 3, 4).
Use a color of 0 to remove the corresponding curve, e.g. col=c(1,2,3,0) to not display the OOB curve.
If col=0 (which gets recycled to c(0,0,0,0)) nothing will be plotted, but plot_gbm will return the number-of-trees at the minima as usual (as described in the Value section below).

ylim

The default ylim="auto" shows more detail around the minima.
Use ylim=NULL for the full vertical range of the curves.
Else specify ylim as usual.

legend.x

The x position of the legend. The default positions the legend automatically.
Use legend.x=NA for no legend.
See the x and y arguments of xy.coords for other options, for example legend.x="topright".

legend.y

The y position of the legend.

legend.cex

The legend cex (the default is 0.8).

grid.col

Default NA. Color of the optional grid, for example grid.col=1.

n.trees

For use by plotres.
The x position of the gray vertical line indicating the n.trees passed by plotres to predict.gbm to calculate the residuals. Plotres defaults to all trees.

col.n.trees

For use by plotres.
Color of the vertical line showing the n.trees argument. Default is "darkgray".

...

Dot arguments are passed internally to plot.default.

Value

This function returns a four-element vector specifying the number of trees at the train, test, CV, and OOB minima respectively.

The minima are calculated after smoothing as specified by this function's smooth argument. By default, only the OOB curve is smoothed. The smoothing algorithm for the OOB curve differs slightly from gbm.perf, so can give a slightly different number of trees.

Note

The OOB curve

The OOB curve is artificially rescaled to force it into the plot. See Chapter 7 in the plotres vignette.

Interaction with plotres

When invoking this function via plotres, prefix any argument of plotres with w1. to tell plotres to pass the argument to this function. For example give w1.ylim=c(0,10) to plotres (plain ylim=c(0,10) in this context gets passed to the residual plots).

Acknowledgments

This function is derived from code in the gbm package authored by Greg Ridgeway and others.

See Also

Chapter 7 in plotres vignette discusses this function.

Examples

if (require(gbm)) {
    n <- 100                            # toy model for quick demo
    x1 <- 3 * runif(n)
    x2 <- 3 * runif(n)
    x3 <- sample(1:4, n, replace=TRUE)
    y <- x1 + x2 + x3 + rnorm(n, 0, .3)
    data <- data.frame(y=y, x1=x1, x2=x2, x3=x3)
    mod <- gbm(y~., data=data, distribution="gaussian",
               n.trees=300, shrinkage=.1, interaction.depth=3,
               train.fraction=.8, verbose=FALSE)

    plot_gbm(mod)

    # plotres(mod)                      # plot residuals

    # plotmo(mod)                       # plot regression surfaces
}