@Override
public SRResult predictSimilarity(List<SRResult> scores) {
if (scores.size() + 1 != simlarityCoefficients.size()) {
throw new IllegalStateException();
}
double weightedScore = simlarityCoefficients.get(0);
for (int i = 0; i < scores.size(); i++) {
double s = scores.get(i) == null ? Double.NaN : scores.get(i).getScore();
if (Double.isNaN(s) || Double.isInfinite(s)) {
s = similarityInterpolator.getInterpolatedScore(i);
}
weightedScore += (s * simlarityCoefficients.get(i + 1));
}
return new SRResult(weightedScore);
}
@Override
public SRResultList predictMostSimilar(
List<SRResultList> scores, int maxResults, TIntSet validIds) {
if (2 * scores.size() + 1 != mostSimilarCoefficients.size()) {
throw new IllegalStateException();
}
TIntSet allIds = new TIntHashSet(); // ids returned by at least one metric
for (SRResultList resultList : scores) {
if (resultList != null) {
for (SRResult result : resultList) {
allIds.add(result.getId());
}
}
}
TIntDoubleHashMap scoreMap = new TIntDoubleHashMap();
for (int id : allIds.toArray()) {
scoreMap.put(id, mostSimilarCoefficients.get(0));
}
int i = 1;
for (SRResultList resultList : scores) {
TIntSet unknownIds = new TIntHashSet(allIds);
double c1 = mostSimilarCoefficients.get(i); // score coeff
double c2 = mostSimilarCoefficients.get(i + 1); // rank coefficient
if (resultList != null) {
for (int j = 0; j < resultList.numDocs(); j++) {
int rank = j + 1;
// expand or contract ranks proportionately
if (validIds != null) {
double k = 1.0 * numTrainingCandidateArticles / validIds.size();
rank = (int) (rank * k);
}
SRResult result = resultList.get(j);
unknownIds.remove(result.getId());
double value = c1 * result.getScore() + c2 * Math.log(rank);
if (debug) {
System.err.format(
"%s %d. %.3f (id=%d), computing %.3f * %.3f + %.3f * (log(%d) = %.3f)\n",
"m" + i, j, value, result.getId(), c1, result.getScore(), c2, rank, Math.log(rank));
}
scoreMap.adjustValue(result.getId(), value);
}
}
// interpolate scores for unknown ids
double value =
c1 * mostSimilarInterpolator.getInterpolatedScore(i / 2)
+ c2 * Math.log(mostSimilarInterpolator.getInterpolatedRank(i / 2));
for (int id : unknownIds.toArray()) {
scoreMap.adjustValue(id, value);
}
i += 2;
}
List<SRResult> resultList = new ArrayList<SRResult>();
for (int id : scoreMap.keys()) {
resultList.add(new SRResult(id, scoreMap.get(id)));
}
Collections.sort(resultList);
Collections.reverse(resultList);
int size = maxResults > resultList.size() ? resultList.size() : maxResults;
SRResultList result = new SRResultList(size);
for (i = 0; i < size; i++) {
result.set(i, resultList.get(i));
}
return result;
}
