protected int gradientUpdate(_Query query) {
double diff;
int i, trainSize = 0;
// Step 1: calculate the ranking score
for (_QUPair pair : query.m_docList) pair.score(m_weight);
m_eval.eval(query);
// Step 2: accumulate the lambdas for each URL
for (_QUPair pair : query.m_docList) {
diff = 0;
if (pair.m_worseURLs != null) {
for (_QUPair worseURL : pair.m_worseURLs) { // force to moving up
diff += Utils.logistic(worseURL.m_score - pair.m_score) * m_eval.delta(pair, worseURL);
trainSize++;
}
}
if (pair.m_betterURLs != null) {
for (_QUPair betterURL : pair.m_betterURLs) { // force to moving down
diff -= Utils.logistic(pair.m_score - betterURL.m_score) * m_eval.delta(betterURL, pair);
trainSize++;
}
}
// Step 3: update weight according to this URL
if (diff != 0) {
for (i = 0; i < pair.m_rankFv.length; i++) m_g[i] -= diff * pair.m_rankFv[i];
}
}
return trainSize;
}
protected void evaluate() {
double r;
m_obj = 0;
m_perf = 0;
m_misorder = 0;
for (_Query query : m_queries) {
// calculate ranking score with latest weight
for (_QUPair pair : query.m_docList) pair.score(m_weight);
if ((r = m_eval.eval(query)) >= 0) // ranking score should already be calculated
m_perf += r;
for (_QUPair pair : query.m_docList) {
if (pair.m_worseURLs != null) {
for (_QUPair worseURL : pair.m_worseURLs) {
if ((r = Utils.logistic(pair.m_score - worseURL.m_score)) > 0) m_obj += Math.log(r);
if (pair.m_score <= worseURL.m_score) m_misorder++;
}
}
if (pair.m_betterURLs != null) {
for (_QUPair betterURL : pair.m_betterURLs) {
if ((r = Utils.logistic(betterURL.m_score - pair.m_score)) > 0) m_obj += Math.log(r);
if (pair.m_score >= betterURL.m_score) m_misorder++;
}
}
}
}
m_misorder /= 2;
}