forestFloor (version 1.11.1)

convolute_ff2: Low-level function to estimate a specific set of feature contributions by corresponding features with kknn-package. Used to estimate goodness-of-fit of surface in show3d.

Description

Low-level function to estimate a selected combination feature contributions as function of selected features with leave-one-out k-nearest neighbor.

Usage

convolute_ff2(ff,
              Xi,
              FCi = NULL,
              k.fun=function() round(sqrt(n.obs)/2),
              userArgs.kknn = alist(kernel="gaussian")            )

Arguments

ff

forestFloor object class "forestFloor_regression" or "forestFloor_multiClass" consisting of at least ff$X and ff$FCmatrix with two matrices of equal size

Xi

integer vector, of column indices of ff$X to estimate by.

FCi

integer vector, column indices of features contributions in ff$FCmatrix to estimate. If more than one , these columns will be summed by samples/rows. If NULL then FCi will match Xi.

k.fun

function to define k-neighbors to consider. n.obs is a constant as number of observations in ff$X. Hereby k neighbors is defined as a function k.fun of n.obs. To set k to a constant use e.g. k.fun = function() 10. k can also be overridden with userArgs.kknn = alist(kernel="Gaussian",kmax=10).

userArgs.kknn

argument list passed to train.kknn function for each convolution, see train.kknn. Arguments in this list have priority of any arguments passed by default by this wrapper function. See argument merger train.kknn

Value

an numeric vector with one estimated feature contribution for any observation

Details

convolute_ff2 is a wrapper of train.kknn to estimate feature contributions by a set of features. This function is e.g. used to estimate the visualized surface layer in show3d function. LOO CV is used to quantify how much of a feature contribution variation can by explained by a given surface. Can in theory also be used to quantify higher dimensional interaction effects, but randomForest do not learn much 3rd order (or higher) interactions. Do not support orderByImportance, thus Xi and FCi points to column order of training matrix X.

Examples

library(forestFloor)
library(randomForest)
library(rgl)
#simulate data
obs=2500
vars = 6 
X = data.frame(replicate(vars,rnorm(obs)))
Y = with(X, X1^2 + 2*sin(X2*pi) + 8 * X3 * X4)
Yerror = 15 * rnorm(obs)
cor(Y,Y+Yerror)^2  #relatively noisy system
Y= Y+Yerror

#grow a forest, remeber to include inbag
rfo=randomForest(X,Y,keep.inbag=TRUE,ntree=1000,sampsize=800)

#obtain 
ff = forestFloor(rfo,X)

#convolute the interacting feature contributions by their feature to understand relationship
fc34_convoluted = convolute_ff2(ff,Xi=3:4,FCi=3:4,  #arguments for the wrapper
                  userArgs.kknn = alist(kernel="gaussian",k=25)) #arguments for train.kknn

#plot the joined convolution
plot3d(ff$X[,3],ff$X[,4],fc34_convoluted,
       main="convolution of two feature contributions by their own vaiables",
       #add some colour gradients to ease visualization
       #box.outliers squese all observations in a 2 std.dev box
       #univariately for a vector or matrix and normalize to [0;1]
       col=rgb(.7*box.outliers(fc34_convoluted), 
               .7*box.outliers(ff$X[,3]),        
               .7*box.outliers(ff$X[,4]))
       )