show3d_new: make 3d-plot of forestFloor topology

In forestFloorStable: (Legacy version 1.5)Visualizes Random Forests with Feature Contributions

Description

2 features features(horizontal XY-plane) and one combined feature contribution (vertical Z-axis). Surface response layer will be estimated(gaussian-kNN by kknn package) and plotted alongside the datapoints. 3D grphic device is rgl.

Usage

show3d_new(ff,
           Xi  = 1:2,
           FCi = NULL,
           col = "#12345678",    
           sortByImportance = TRUE,
           surface=TRUE,   
           combineFC = sum,  
           zoom=1.2,       
           grid.lines=30,  
           limit=3, 
           kknnGrid.args = alist(),  
           plot.rgl.args = alist(),  
           surf.rgl.args = alist()   
           )

Arguments

ff	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	colour vector points colour or colour palette, can also be passed as promise in plot.rgl.args
sortByImportance	booleen should indices count 'variable importance' order or matrix/data.frame order
surface	should a surface be plotted also
combineFC	how should feature contributions be combined
zoom	#grid can be expanded in all directions by a factor ,zoom
grid.lines	#how many grid lines should be used
limit	#sizing of grid does not concider outliers, outside limit of e.g. 3 sd deviations univariately
kknnGrid.args	argument list, any possiple 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 paremters. k is set by default automatically to a half times the square root of observations, which often gives a reasonable balance between robustness and apdeptness. 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 transparancy 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.

Details

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 of forestFloor class, and should at least contain $X the data.frame of training data, $FCmatrix the feature contributions matrix.

This wrapper function will first plot points. Any graphical argument of the 3D plotting scene can be forwarded through plot.rgl.args, see plot3drgl. The transparency of points could e.g. be lowered from 0.4 to 0.1 and the size increased by using plot.rgl.args = alist(alpha=0.1,size=5)
Secondly, a grid response surface layer will be estimated(gaussian-kNN), on basis of selected features and feature contributions. Any argument can passed to kknnkknn by kknnGrid.args, e.g. kknnGrid.args = alist(k=5, kernel="triangular")
Thirdly the surface grid is plotted. Any argument can be forwarded to persp3drgl through surf.rgl.args. The surface layer could be coloured and made more transparent by passing surf.rgl.args = alist(alpha=0.2, col=c("red","green","blue")). Notice some graphical settings, should as axis labels(xlab) must be defined by plot.rgl.args.

policy of merging user arguments with wrapper arguments is handled by append.overwrite.alistsforestFloor

Value

no value

Author(s)

Soren Havelund Welling

Examples

## Not run: 
rm(list=ls())
library(forestFloorStable)
#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_new(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_new(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_new(ff,1:2,col=Col,plot.rgl=list(size=5))


## End(Not run)

forestFloorStable documentation built on May 2, 2019, 5:22 p.m.