void AdaRanker::learn(const QueryData &data, const Metric &metric)
{
weakRankerWeights = std::vector<double>(data.nfeature(), 0.0);
std::vector<int> weakRankerChecked(data.nfeature(), 0);
// precompute weak rankers
std::vector<std::vector<double> > weakRankerScores(data.nquery(), std::vector<double>(data.nfeature(), 0.0));
for(int q = 0; q < data.nquery(); q++)
{
for(int f = 0; f < data.nfeature(); f++)
{
weakRankerScores[q][f] = metric.measure(weakRank(data.getQuery(q), f));
}
}
// a log of ranker scores
std::vector<double> scores(1, metric.measure(rank(data)));
// iterate over each feature
std::vector<double> queryWeights(data.nquery(), 1.0/data.nquery());
for(int t = 0; t < data.nfeature(); t++)
{
// compute the query weighted score of unchecked each weak ranker
std::vector<double> weightedScores(data.nfeature(), 0);
for(int f = 0; f < data.nfeature(); f++)
{
if(weakRankerChecked[f] == 0)
{
for(int q = 0; q < data.nquery(); q++)
{
weightedScores[f] += weakRankerScores[q][f]*queryWeights[q];
}
}
}
// find best weak ranker
int bestWR = 0;
for(int f = 0; f < data.nfeature(); f++)
{
if(weightedScores[f] > weightedScores[bestWR])
{
bestWR = f;
}
}
// compute alpha
double num = 0.0;
double den = 0.0;
for(int q = 0; q < data.nquery(); q++)
{
num += queryWeights[q]*(1.0 + weakRankerScores[q][bestWR]);
den += queryWeights[q]*(1.0 - weakRankerScores[q][bestWR]);
}
double alpha = 0.5*(log(num/den));
weakRankerWeights[bestWR] = alpha;
weakRankerChecked[bestWR] = 1;
// only update query weights if the new weak ranker improves overall performance
double newScore = metric.measure(rank(data));
if((newScore-scores.back()) > 1e-4)
{
scores.push_back(newScore);
std::vector<double> queryProbs;
for(int q = 0; q < data.nquery(); q++)
{
queryProbs.push_back(exp(-metric.measure(rank(data.getQuery(q)))));
}
double totalQueryProb = std::accumulate(queryProbs.begin(), queryProbs.end(), 0.0);
for(int q = 0; q < data.nquery(); q++)
{
queryWeights[q] = queryProbs[q]/totalQueryProb;
}
}
else
{
weakRankerWeights[bestWR] = 0.0;
}
}
}
