plot.gg_partial_rfsrc: categorical partial dependence stacks per-observation predictions instead of averaging

[ehrlinger]

Observed

For a factor or 0/1 predictor, plot(gg_partial_rfsrc(...)) draws bars whose height is the sum of every observation's prediction, so a class probability reads in the hundreds.

Reproduce (ggRandomForests 4.0.0, randomForestSRC 3.9.0)

library(randomForestSRC); library(ggRandomForests)
set.seed(1); n <- 400
d <- data.frame(x = rnorm(n), g = rbinom(n, 1, .4))
d$y <- factor(rbinom(n, 1, plogis(-1 + d$x + d$g)))
p <- gg_partial_rfsrc(rfsrc(y ~ x + g, d, ntree = 100), xvar.names = c("x", "g"))
nrow(p$continuous)    # 25: averaged over the grid
nrow(p$categorical)   # 800: one prediction per observation per level, not averaged
tapply(p$categorical$yhat, p$categorical$x, mean)  # 0.626, 0.482
b <- ggplot2::ggplot_build(plot(p))  # categorical panel is geom_bar; y accumulates 1.00, 1.84, 2.81, ...

The continuous component is a partial-dependence curve (averaged); the categorical component is individual conditional expectations, and geom_bar stacks them.

Expected

The categorical panel shows the mean prediction per level (0.63 and 0.48 here), as the continuous panel does, or a boxplot of the per-observation predictions if the spread is the point. Either way the y-axis should be on the response scale.

Seen in hvtiRtemplates' rfs-explain and rfc-explain reports: bars near 1800 (survival) and 600 (classification) for 0/1 predictors.

🤖 Generated with Claude Code