show3d: make forestFloor 3D-plot of random forest feature...
In sorhawell/forestFloor: Visualizes Random Forests with Feature Contributions

Description Usage Arguments Details Value Author(s) Examples

2 features features(horizontal XY-plane) and one combined feature contribution (vertical Z-axis). Surface response layer will be estimated(kknn package) and plotted alongside the data points. 3D graphic device is rgl. Will dispatch methods show3d.forestFloor_regression for regression and show3d_forestFloor_multiClass for classification.

## S3 method for class 'forestFloor_regression'
 show3d(
      x,
      Xi  = 1:2,
      FCi = NULL,
      col = "#12345678",
      plotTest = NULL,
      orderByImportance = TRUE,
      surface=TRUE,   
      combineFC = sum,  
      zoom=1.2,       
      grid.lines=30,  
      limit=3, 
      cropPointsOutSideLimit = TRUE,
      kknnGrid.args = alist(),  
      plot.rgl.args = alist(),  
      surf.rgl.args = alist(),
      user.gof.args = alist(),
      plot_GOF = TRUE,
      ...)

## S3 method for class 'forestFloor_multiClass'
show3d(
      x,
      Xi,
      FCi=NULL,
      plotTest = NULL,
      label.seq=NULL,
      kknnGrid.args=list(NULL),
      plot.rgl.args=list(),
      plot_GOF=FALSE,
      user.gof.args=list(NULL),
      ...)

`x`	forestFloor" class object
`Xi`	integer vector of length 2 indices of feature columns
`FCi`	integer vector of length 1 to p variables indices of feature contributions columns
`col`	a colour vector. One colour or colour palette(vector).
`plotTest`	NULL(plot by test set if available), TRUE(plot by test set), FALSE(plot by train), "andTrain"(plot by both test and train)
`orderByImportance`	should indices order by 'variable importance' or by matrix/data.frame order?
`surface`	should a surface be plotted also?
`combineFC`	a row function applied on selected columns(FCi) on $FCmatrix or $FCarray. How should feature contributions be combined? Default is `sum`.
`zoom`	grid can be expanded in all directions by a factor
`grid.lines`	how many grid lines should be used. Total surface anchor points in plot is grid.lines^2. May run slow above 200-500 depending on hardware.
`limit`	a number. Sizing of grid does not consider outliers outside this limit of e.g. 3 SD deviations univariately.
`cropPointsOutSideLimit`	#if points exceed standard deviation limit, they will not be plotted
`kknnGrid.args`	argument list, any possible arguments to kknnkknn These default wrapper arguments can hereby be overwritten: wrapper = alist( formula=fc~., # do not change train=Data, # do not change k=k, # integer < n_observations. k>100 may run slow. kernel="gaussian", #distance kernel, other is e.g. kernel="triangular" test=gridX #do not change ) see kknnkknn to understand parameters. k is set by default automatically to a half times the square root of observations, which often gives a reasonable balance between robustness and adeptness. k neighbors and distance kernel can be changed be passing kknnGrid.args = alist(k=5,kernel="triangular",scale=FALSE), hereby will default k and default kernel be overwritten. Moreover the scale argument was not specified by this wrapper and therefore not conflicting, the argument is simply appended.
`plot.rgl.args`	pass argument to rgl::plot3d, can override any argument of this wrapper, defines plotting space and plot points. See plot3d for documentation of graphical arguments. wrapper_arg = alist( x=xaxis, #do not change, x coordinates y=yaxis, #do not change, y coordinates z=zaxis, #do not change, z coordinates col=col, #colouring evaluated within this wrapper function xlab=names(X)[1], #xlab, label for x axis ylab=names(X)[2], #ylab, label for y axis zlab=paste(names(X[,FCi]),collapse=" - "), #zlab, label for z axis alpha=.4, #points transparency size=3, #point size scale=.7, #z axis scaling avoidFreeType = T, #disable freeType=T plug-in. (Postscript labels) add=FALSE #do not change, should graphics be added to other rgl-plot? )
`surf.rgl.args`	wrapper_arg = alist( x=unique(grid[,2]), #do not change, values of x-axis y=unique(grid[,3]), #do not change, values of y-axis z=grid[,1], #do not change, response surface values add=TRUE, #do not change, surface added to plotted points alpha=0.4 #transparency of surface, [0;1] ) see rgl::persp3d for other graphical arguments notice the surface is added onto plotting of points, thus can e.g. labels not be changed from here.
`label.seq`	a numeric vector describing which classes and in what sequence to plot. NULL is all classes ordered is in levels in x$Y of forestFloor_mulitClass object x.
`user.gof.args`	argument list passed to internal function convolute_ff2, which can modify how goodness-of-fit is computed. Number of neighbors and kernel can be set manually with e.g. list(kmax=40,kernel="gaussian"). Default pars should work already in most cases. Function convolute_ff2 computed leave-one-out CV prediction the feature contributions from the chosen context of the visualization.
`plot_GOF`	Boolean TRUE/FALSE. Should the goodness of fit be computed and plotted is main of 3D plot? If false, no GOF input pars are useful.
`...`	not used at the moment

show3d plot one or more combined feature contributions in the context of two features with points representing each data point. The input object must be a "forestFloor_regression" or "forestFloor_multiClass" S3 class object , and should at least contain $X the data.frame of training data, $FCmatrix the feature contributions matrix. Usually this object are formed with the function forestFloor having a random forest model fit as input. Actual visualization differs for each class.

no value

Soren Havelund Welling

## Not run: 
## avoid testing of rgl 3D plot on headless non-windows OS
## users can disregard this sentence.
if(!interactive() && Sys.info()["sysname"]!="Windows") skipRGL=TRUE

library(forestFloor)
library(randomForest)
#simulate data
obs=2500
vars = 6 

X = data.frame(replicate(vars,rnorm(obs)))
Y = with(X, X1^2 + sin(X2*pi) + 2 * X3 * X4 + 1 * rnorm(obs))


#grow a forest, remeber to include inbag
rfo=randomForest(X,Y,keep.inbag = TRUE,sampsize=1500,ntree=500)

#compute topology
ff = forestFloor(rfo,X)


#print forestFloor
print(ff) 

#plot partial functions of most important variables first
plot(ff) 

#Non interacting functions are well displayed, whereas X3 and X4 are not
#by applying different colourgradient, interactions reveal themself 
Col = fcol(ff,3)
plot(ff,col=Col) 

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

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

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

## End(Not run)