I would recommend using MDS since it gives the best summary at a glance.
But in favor of interpretability, after looking at variable importances, you could pick the k
most significant variables (if you have a large number of predictors) and use classCenter
function on the proximity matrix. This will let you visualize the representative point for each class, for every variable.
Example code:
library(ggplot2)
library(reshape2)
data(iris)
iris.rf <- randomForest(iris[,-5], iris[,5], prox=TRUE)
iris.p <- melt(classCenter(iris[,-5], iris[,5], iris.rf$prox), id=rownames())
iris.m <- melt(iris, id='Species', variable.name = 'Var', value.name = 'Measurement')
names(iris.p) <- names(iris.m)
ggplot(iris.m, aes(x=Species, color=Species)) +
geom_point(aes(y=Measurement)) +
geom_point(data=iris.p,aes(y=Measurement), size=4, pch=15, color='grey40') +
facet_grid(.~Var)
Output:

If you have new points with Species prediction, simply add another geom_point
layer with the new data and change the color
or pch
attributes to differentiate from old points. This will let you visualize the new points relative to training data as well as class representatives from training data.
Alternatively, you could take the variables pairwise and see the class prototypes/ centers for each variable combination. The help file ?classCenter
gives an example of that.