void KNN::train(Examples& exs){
TRACE_V(TAG,"train");
//Maybe we didnt calculate this before...
stats->calculateIDF();
for(int i = 0; i < exs.getNumberOfNumericalAttibutes(); i++){
maxv[i] = numeric_limits<double>::min();
minv[i] = numeric_limits<double>::max();
}
for(ExampleIterator e = exs.getBegin(); e != exs.getEnd(); e++){
vector<string> textTokens = (e)->getTextTokens();
vector<int> textFrequencyTokens = (e)->getTextFrequency();
string exampleClass = (e)->getClass();
string eId = (e)->getId();
double docSize = 0.0;
// cout<<" Tokens categoricos = " << tokens.size() << endl;
for(unsigned int i = 3; i < textTokens.size(); i++){
int tf = textFrequencyTokens[i-3];
string termId = textTokens[i];
double tfidf = tf * stats->getIDF(termId);
docSize += (tfidf * tfidf);
docWeighted dw(eId, tfidf);
termDocWset[termId].insert(dw);
}
vector<double> numTokens = (e)->getNumericalTokens();
for(unsigned int i = 0; i < numTokens.size(); i++){
if(greaterThan(numTokens[i], maxv[i])){
maxv[i] = numTokens[i];
}
if(lesserThan(numTokens[i], minv[i])){
minv[i] = numTokens[i];
}
}
exNumTrain[eId] = numTokens;
exCatTrain[eId] = (e)->getCategoricalTokens();
docTrainSizes[eId] = docSize;
}
}
void KNN::test(Examples& exs){
TRACE_V(TAG,"test");
//Statistics:
map<string,unsigned long long> classHits;
map<string,unsigned long long> classMiss;
map<string,unsigned long long> mappedDocs;
map<string,unsigned long long> docsPerClass;
int numExamples = 0;
for(ExampleIterator it = exs.getBegin(); it != exs.getEnd(); it++){
numExamples++;
if(numExamples % 100 == 0)
cout<<"Evaluated: " << numExamples<<endl;
Example ex = *it;
vector<string> textTokens = ex.getTextTokens();
vector<int> textFreqTokens = ex.getTextFrequency();
vector<double> numTokens = ex.getNumericalTokens();
vector<string> catTokens = ex.getCategoricalTokens();
string eId = ex.getId();
string classId = ex.getClass();
map<string, double> examplesTestSize;
//credibility to each class
if((usingKNNOptimize && !valuesSaved ) || !usingKNNOptimize){
for(unsigned int i = 3; i < textTokens.size(); i++){
string termId = textTokens[i];
int tf = textFreqTokens[i-3];
for(set<string>::iterator classIt = stats->getClasses().begin(); classIt != stats->getClasses().end(); classIt++) {
double tfidf = tf * getContentCredibility(termId, *classIt);
examplesTestSize[*classIt] += (tfidf * tfidf);
}
}
}
map<string, double> similarity;
if(usingKNNOptimize && valuesSaved){
similarity = saveValues[eId];
}
else{
for(unsigned int i = 3; i < textTokens.size();i++){
string termId = textTokens[i];
int tf = textFreqTokens[1-3];
for(set<docWeighted, docWeightedCmp>::iterator termIt = termDocWset[termId].begin(); termIt != termDocWset[termId].end(); termIt++){
string trainClass = stats-> getTrainClass(termIt->docId);
double trainDocSize = docTrainSizes[termIt->docId];
double trainTermWeight = termIt->weight;
double testTermWeight = tf * getContentCredibility(termId, trainClass);
similarity[termIt->docId] += ( - ( trainTermWeight / sqrt(trainDocSize) * testTermWeight / sqrt(examplesTestSize[trainClass]) ));
// cout<<"sim = " << similarity[termIt->docId] <<endl;
}
}
//numerical KNN
for(map<string, vector<double> >::iterator trainIt = exNumTrain.begin(); trainIt != exNumTrain.end(); trainIt++){
double dist = 0.0;
for(unsigned int i = 0; i < numTokens.size(); i++){
double a = minMaxNorm(numTokens[i],i);
double b = minMaxNorm(exNumTrain[trainIt->first][i],i);
double val = (a-b)*(a-b);
//double val = (numTokens[i] - exNumTrain[trainIt->first][i]) * ( numTokens[i] - exNumTrain[trainIt->first][i]);
// cout<<numTokens[i] << " - " <<exNumTrain[trainIt->first][i] <<endl;
// cout<<"a = " << a << " b = " << b << " val =" << val<<endl;
if( greaterThan(dist + val, numeric_limits<double>::max())){
// cerr<<"OOOOOOOOOOOOOOOPA!!!"<<endl;
// exit(0);
dist = numeric_limits<double>::max() - 1.0;
break;
}
dist += val;
// cout<<"dist =" << dist<<endl;
}
similarity[trainIt->first] += dist;
}
//categorical KNN
for(map<string, vector<string> >::iterator trainIt = exCatTrain.begin(); trainIt != exCatTrain.end(); trainIt++){
double dist = 0.0;
for(unsigned int i = 0; i < catTokens.size(); i++){
string trainTok = exCatTrain[trainIt->first][i];
string testTok = catTokens[i];
double catCred = getCategoricalCredibility(i, testTok, stats->getTrainClass(trainIt->first));
// cout<<"catCred = " <<catCred<<endl;
// cout<<" i = " << i << "teste = " << testTok<<" treino = " << trainTok<<endl;
if(trainTok != testTok){
// dist+= 1.0/(catCred+ 1.0) + 1.0;
dist+= 1.0/(catCred+ 1.0);
// cout<<"dist = " << dist<<endl;
}
}
similarity[trainIt->first] += dist;
}
// cout<<"class = " << classId << " doc = " << trainIt->first<< " docClass = " << stats->getTrainClass(trainIt->first) << " dist="<<dist<< " 1/dist = " <<1.0/dist<< " sqrt = "<<sqrt(dist)<<endl;
}
if(!valuesSaved && usingKNNOptimize){
saveValues[eId] = similarity;
}
//sim of each example in test set
set<docWeighted, docWeightedCmp> sim;
for(map<string, double>::iterator testIt = similarity.begin(); testIt != similarity.end(); testIt++){
//calculating graph credibility....if so
vector<double> graphsCreds(graphsCredibility.size());
double similarityValue = testIt->second;
// cout<< " eid = " << eId << " eclass = " << classId << " traindocclass = " << stats->getTrainClass(testIt->first) << " similarit = " << similarityValue<< endl;
for(unsigned int g = 0 ; g < graphsCredibility.size(); g++){
double gsim = getGraphCredibility(g, eId, stats->getTrainClass(testIt->first));
similarityValue /= (0.5+gsim);
}
//never change this, it is necessary
docWeighted dw(testIt->first, similarityValue);
sim.insert(dw);
}
string predictedLabel = getPredictedClass(sim);
computeConfusionMatrix(classId, predictedLabel);
// if(io->usingPredictionsFile)
savePrediction(eId, classId, predictedLabel);
if(predictedLabel == classId){
classHits[classId] ++;
}
else{
classMiss[classId]++;
}
mappedDocs[predictedLabel]++;
docsPerClass[classId]++;
}
if(valuesSaved == false){
valuesSaved = true;
}
calculateF1(classHits,classMiss,docsPerClass, mappedDocs);
// showConfusionMatrix();
}