function [ ypred ] = NicForest_apply( f_out, xtest )
% [ ypred ] = NicForest_apply( forest, xtrain, xtest , ytest) applies a
% model in forests, originally trained on xtrain, to xtest. ytest is used
% to determine if it is a binary or regression problem.
% $LastChangedBy: alistair $
% $LastChangedDate: 2012-05-16 13:48:37 +0100 (Wed, 16 May 2012) $
% $Revision: 10 $
% Originally written on MACI64 by Louis Mayaud, 25-April-2012 14:05:26
% Contact: alistairewj@gmail.com
forests = f_out.forests;
xtrain = f_out.xtrain;
NaNNbre = f_out.TrainNaN; % number of non-nans in training data set
N = size(xtest,1);
idxTraining = [false(size(xtest,1),1);true(size(xtrain,1),1)]; % elements used for ranking
xtest_rk = tiedrankrelative([xtest;xtrain],idxTraining);
xtest_rk = xtest_rk(1:size(xtest,1),:); % remove training set
%=== NORMALIZE DATA ===%
%=== Scale ranks between 0->1
xtest_rk_normalize = bsxfun(@rdivide,xtest_rk,NaNNbre);
xtest_rk_normalize = norminv(xtest_rk_normalize,0,1);
[NForests , ~ , Ntrees] = size(forests);
Ntrees = Ntrees -1 ; % remove the intercept
% create prediction vector for each forest
ypred=zeros(size(xtest_rk,1),NForests);
% For each forest
for i=1:NForests
ypred(:,i)=ypred(:,i)+forests(i,1,Ntrees+1);
val = zeros(N,1);
for j=1:Ntrees % For each tree
val= val + apply_tree( forests(i,:,j) , xtest_rk , NaNNbre );
end
ypred(:,i)=ypred(:,i)+val;
end
%=== invlogit if binomial case
%=== method: regression or classification
if strcmp(f_out.Family,'binomial')
ypred = invlogit(ypred);
%=== sum across forests and normalize by number of forests
ypred = sum(ypred,2)/NForests;
else
% un-normalize predictions using normcdf
ypred = sum(ypred,2)/NForests;
%=== re-scale to original values
ypred = normcdf(ypred,0,1)*f_out.ynum;
%=== use inverse ranking mapping to find actual predicted values
%=== first get the ranks of the predictions/training values
yrk = [f_out.yrk;ypred];
idxPred = [false(size(f_out.yrk));true(size(ypred))];
[yrk,idxSort] = sort(yrk,1,'ascend');
idxPred = idxPred(idxSort);
%=== now order the actual y values in the training set similarly
ytrain = zeros(numel(yrk),1);
ytrain(~idxPred) = f_out.ytrain(idxSort(~idxPred));
%=== now calculate predictions using weighted average
ypred_sorted = propogateValues(ytrain,yrk);
idxUnsort = 1:numel(idxSort);
idxUnsort(idxSort) = idxUnsort;
ypred_sorted = ypred_sorted(idxUnsort);
ypred = ypred_sorted(idxPred(idxUnsort));
end
end
function [x] = invlogit(x)
x = 1./(1+exp(-x));
end
function [ x ] = propogateValues(x,y)
%PROPOGATEVALUES Carry non-zero values forward in vector
% [ x ] = propogateValues(x)
%
%
% Inputs:
% x - A vector containing 0s and non-zeroes, in which the zeroes
% should be replaced with the first prior non-zero value.
% y - A vector containing ranks of x, with no 0s.
%
% Outputs:
% x - A vector with 0s replaced by the first prior non-zero value.
%
%
% Example
% x = [0,1,0,0,6,0,0,0,5,0]
% [ x ] = propogateValues(x)
%
% See also FIND DIFF
% Copyright 2012 Alistair Johnson
% $LastChangedBy: alistair $
% $LastChangedDate: 2012-05-16 13:48:37 +0100 (Wed, 16 May 2012) $
% $Revision: 10 $
% Originally written on GLNXA64 by Alistair Johnson, 08-May-2012 11:27:49
% Contact: alistairewj@gmail.com
valind = find(x==0,1,'first');
%=== handle the extrenum
if valind(1) == 1
end
valind = find(x==0,1,'last');
if valind(end) == numel(x)
end
idx0 = x==0;
x1 = x; x2 = flipud(x);
valind1 = find(x1);
valind2 = find(x2);
%=== x1 = propogate values forward
x1(valind1(2:end)) = diff(x1(valind1));
x1 = cumsum(x1);
%=== x2 = propogate values backward
x2(valind2(2:end)) = diff(x2(valind2));
x2 = cumsum(x2);
x2=flipud(x2);
%=== y1/y2 are the same but using the ranks
y1=y; y1(idx0) = 0;
y2=y; y2(idx0) = 0; y2=flipud(y2);
y1(valind1(2:end)) = diff(y1(valind1));
y1 = cumsum(y1);
y2(valind2(2:end)) = diff(y2(valind2));
y2 = cumsum(y2);
y2=flipud(y2);
%=== calculate weights using y values which are ranks
y1=y1(idx0); y2=y2(idx0); x1=x1(idx0); x2=x2(idx0);
y0 = y(idx0);
x(idx0) = ((y0-y1).*x2 + (y2-y0).*x1)./(y2-y1);
end