tests/stdUsage.R

In forestFloor: Visualizes Random Forests with Feature Contributions

if(!interactive()) Sys.setenv(RGL_USE_NULL=TRUE) #disable RGL for headless machines
library(forestFloor)
library(randomForest)
#simulate data
obs=2000
vars = 6 

X = data.frame(replicate(vars,rnorm(obs)))
Xtest = data.frame(replicate(vars,rnorm(obs)*3))
Y = with(X, X1^2 + sin(X2*pi) + 2 * X3 * X4 + .5 * rnorm(obs))


#grow a forest, remeber to include inbag
rf41=randomForest(X,Y,keep.inbag = TRUE,sampsize=499,ntree=100,importance = F)
#compute feature contributions
out = tryCatch({ff41 = forestFloor(rf41,X,bootstrapFC = TRUE,impType=1)},warning = function(w) w)
if(out$message != "found nothing of importance, revert to fallback") stop("wrong warning")

#if(colnames(ff41$importance) != "%IncMSE") stop("wrong imp")
ff41 = forestFloor(rf41,X,bootstrapFC = TRUE,impType=2)
if(colnames(ff41$importance) != "IncNodePurity") stop("wrong imp")
ff41 = forestFloor(rf41,X,bootstrapFC = TRUE,impType=NULL)
if(colnames(ff41$importance) != "IncNodePurity") stop("wrong imp")

#grow a forest, remeber to include inbag
rf42=randomForest(X,Y,keep.inbag = TRUE,sampsize=499,ntree=100,importance = T)
#compute feature contributions
ff42 = forestFloor(rf42,X,bootstrapFC = TRUE,impType=1)
if(colnames(ff42$importance) != "%IncMSE") stop("wrong imp")
ff42 = forestFloor(rf42,X,bootstrapFC = TRUE,impType=2)
if(colnames(ff42$importance) != "IncNodePurity") stop("wrong imp")
ff42 = forestFloor(rf42,X,bootstrapFC = TRUE,impType=NULL)
if(colnames(ff42$importance) != "%IncMSE") stop("wrong imp")




#test accuracy of feature contributions
#y_hat_OOB = row sum FC + Y_grandMean
FC.residuals = rf42$predicted - apply(ff42$FCmatrix[ff42$isTrain,],1,sum) - mean(Y)
if(max(abs(FC.residuals))>1E-12) stop(
  paste0("When testing if:  y_hat_OOB = row sum FCmatrix + Y_grandMean
  one/some FCs error exceed allowed 1e-12, found.error=",max(abs(FC.residuals)))
)

#test same results are reached with Xtest
ff43 = forestFloor(rf42,X,Xtest,bootstrapFC = TRUE)
if(max(abs(ff43$FCmatrix[ff43$isTrain,]-ff42$FCmatrix)) > 1E-12) stop(
  "forestFloor with/without Xtest gives different feature contributions"
)


#print forestFloor
print(ff42) 

#plot partial functions of most important variables first
plot(ff42,orderByImportance=TRUE) 
plot(ff43,orderByImportance=TRUE)

#Non interacting functions are well displayed, whereas X3 and X4 are not
#by applying different colourgradient, interactions reveal themself 
#also a k-nearest neighbor fit is applied to evaluate goodness of fit
Col=fcol(ff43,3,orderByImportance=FALSE)
plot(ff43,col=Col,plot_GOF=TRUE,speed=T) 


##make test set grey tone to show if point of test is extrapolated
Col=fcol(ff43,3,orderByImportance=FALSE,plotTest="andTrain",alpha=.2)
Col[ff43$isTrain] = "#000000FF"
plot(ff43,col=Col,speed=T,plotTest="andTrain",plot_GOF=F) 


#if ever needed, k-nearest neighbor parameters for goodness-of-fit can be access through convolute_ff
#a new fit will be calculated and added to forstFloor object as ff42$FCfit
ff43 = convolute_ff(ff43,userArgs.kknn=alist(kernel="epanechnikov",kmax=5))
plot(ff43,col=Col,plot_GOF=TRUE)

#in 3D the interaction between X3 and X reveals itself completely
show3d(ff43,3:4,col=Col,plot.rgl=list(size=5),orderByImportance=FALSE)
Col=fcol(ff43,1:2,orderByImportance=FALSE)
show3d(ff43,1:2,col=Col,plot.rgl=list(size=5),orderByImportance=FALSE)


Col=fcol(ff43,1:2,plotTest="andTrain",orderByImportance=FALSE)
show3d(ff43,1:2,col=Col,plot.rgl=list(size=5),orderByImportance=FALSE)



#although no interaction, a joined additive effect of X1 and X2
#colour by FC-component FC1 and FC2 summed
Col = fcol(ff43,1:2,orderByImportance=FALSE,X.m=FALSE,RGB=TRUE,plotTest = "a")
plot(ff43,col=Col,plotTest = "a") 
show3d(ff43,1:2,col=Col,plot.rgl=list(size=5),orderByImportance=FALSE,plotTest = "a") 

#...or two-way gradient is formed from FC-component X1 and X2.
Col = fcol(ff42,1:2,orderByImportance=FALSE,X.matrix=TRUE,alpha=0.8) 
plot(ff42,col=Col) 
show3d(ff42,1:2,col=Col,plot.rgl=list(size=5),orderByImportance=FALSE,plot_GOF=T)


#testing single feature contribution indice plot, implemented from 1.8.7
show3d(ff42,Xi=1:2,
       FCi=1, #only one feature contribution is chosen
       col=Col,plot.rgl=list(size=5),
       orderByImportance=FALSE,plot_GOF=T)