function results=Experiments(dataset,frac,probLabeled,algo,lambda,lambda_u,multiple_switches)
%  Reproduces Experimental results of the paper:
%  "Large Scale Semi-supervised Linear SVMs"
%   V. Sindhwani & S.S.Keerthi
%   Book Chapter in Large Scale Kernel Machines,
%   Eds: Bottou, Decoste, Chapelle,Weston. 2006
%   A version of this paper appeared in ACM SIGIR 2006.
% 
%
% Inputs:
%           
%        dataset: 'aut-avn','real-sim','ccat','gcat','33-36','pcmac'
%           frac: specifies what fraction of the full data to be used as
%                 training (labeled & unlabeled). 
%                 1-frac is the fraction used as test examples.
%    probLabeled: specifies what fraction of training data to be used as
%                 labeled. Can be a vector. 
%           algo: algorithm to use: (0: rls, 1:l2svm 2:tsvm 3:da)
%         lambda: regularization parameter (default 0.001 was used for all
%                 experiments)
%       lambda_u: regularization parameter (used 1 for all datasets except
%                 aut-avn, ccat where 10 was used.
% multiple_switches: for TSVM, how many switches to use (set 1 for single
%                    switching, Inf (u/2) for maximum switching)
% Use the following values -- 
%
% aut-avn, real-sim:  
%       frac=0.5, probLabeled=[0.00125 0.0025 0.0050 0.0100 0.0200 0.0400];
% ccat, gcat:
%       frac=0.75, probLabeled=[0.0025 0.0050 0.0100 0.0200 0.0400 0.0800]; 
% 33-36: frac=0.5, probLabeled=[0.00125 0.0025 0.0050 0.0100 0.0200 0.0400]; 
% pcmac: frac=0.75,probLabeled=[0.0250 0.0500 0.0750 0.1000 0.1250 0.1500];
%

seed=1e5;
results.seed=seed;
results.algo=algo;
results.Dataset=dataset;
results.probLabeled=probLabeled;
results.numLabeled=zeros(length(probLabeled),1);

load(dataset);
r=sum(Y>0)/length(Y);

if exist('multiple_switches','var')
   if isinf(multiple_switches)
       S=length(Y);
   else
       S=multiple_switches;
   end
end
cmdline=[' -A ' num2str(algo) ' -W ' num2str(lambda) ' -U ' num2str(lambda_u) ' -R ' num2str(r) ' -S ' num2str(S)]; 

dataset

cmdline

[pos,neg]=make_splits(Y,seed);

n=ceil(frac*length(Y));
results.TrainSize=n;
results.TestSize=length(Y)-n;
np=ceil(r*n);
nn=n-np;

numLabeled=ceil(probLabeled*n);
results.numLabeled=numLabeled;

for i = 1:size(pos,1)
for j = 1:length(probLabeled)
    
l=numLabeled(j);
lp=ceil(l*r);
ln=l-lp;

L=[pos(i,1:lp) neg(i,1:ln)];
U=[pos(i,lp+1:np) neg(i,ln+1:nn)];
T=[pos(i,np+1:end) neg(i,nn+1:end)];
 
Ytrain=zeros(length([L U]),1);
Ytrain(1:length(L))=Y(L);

[w,o,f]=svmlin(cmdline,X([L U]',:),Ytrain);

results.time(i,j)=f;
error_u(i,j)=performance(o(l+1:end),Y(U));

[w,o]=SVMlin([],X(T',:),[],w);
error_t(i,j)=performance(o,Y(T));


fprintf(1,'transduction error (split %d , lab %d ) = %f (time %f)\n',i,l, error_u(i,j),f); 
fprintf(1,'        test error (split %d , lab %d ) = %f\n',i,l, error_t(i,j)); 

end
end

results.Unlab.error=error_u;
results.Test.error=error_t;

function [pos,neg]=make_splits(Y,seed)
idx_pos=find(Y>0); idx_pos=idx_pos';
idx_neg=find(Y<0); idx_neg=idx_neg';
rand('seed',seed);
pos=zeros(10,length(idx_pos));
neg=zeros(10,length(idx_neg));

for i=1:10 
  pos(i,:)=idx_pos(randperm(length(idx_pos)));
  neg(i,:)=idx_neg(randperm(length(idx_neg)));
end

function error=performance(f,y)
    guess=sign(f);
    error=1-sum(guess==y)/length(y);