function Y_hat = kernel_calibrate( X_new, X, Y, besth )
% Given a set of training data decision values (X) and their labels (Y),
% use kernel regression to estimate outcome probabilities (Y_hat) for a
% new set of data decision values (X_new)
%
% Inputs:
% X_new - new data (1-dimensional) points, usually classifier outputs
% X - training data points
% Y - training data labels (0/1)
%
% Y_hat - estimated outcome probabilities for each of the new points
% make sure negative labels aren't -1
Y(Y < 0) = 0;
%{
% test over a range of kernel bandwidths using inner cross-validation
% use hosmer-lemeshow chi-squared score as the selection criteria
[TRN,TST] = cv_split(5,Y);
hs = 10.^(-3:0.1:0);
bestH = Inf;
besth = NaN;
for h=hs
meanH = 0;
for iter=1:length(TRN)
icv_Y_hat = kernel_regress(X(TST{iter}), X(TRN{iter}), Y(TRN{iter}), h);
H = lemeshow([Y(TST{iter})' icv_Y_hat], 0);
meanH = meanH + H;
end
meanH = meanH / length(TRN);
%disp(['h: ' num2str(h) ' H: ' num2str(meanH)]);
if meanH < bestH
bestH = meanH;
besth = h;
end
end
%}
Y_hat = kernel_regress(X_new, X, Y, besth);
end