Preconditions.checkArgument(at >= 1, "at must be at least 1");
Preconditions.checkArgument(evaluationPercentage > 0.0 && evaluationPercentage <= 1.0,
"Invalid evaluationPercentage: %s", evaluationPercentage);
int numItems = dataModel.getNumItems();
RunningAverage precision = new FullRunningAverage();
RunningAverage recall = new FullRunningAverage();
RunningAverage fallOut = new FullRunningAverage();
RunningAverage nDCG = new FullRunningAverage();
LongPrimitiveIterator it = dataModel.getUserIDs();
while (it.hasNext()) {
long userID = it.nextLong();
if (random.nextDouble() >= evaluationPercentage) {
// Skipped
continue;
}
long start = System.currentTimeMillis();
PreferenceArray prefs = dataModel.getPreferencesFromUser(userID);
int size = prefs.length();
if (size < 2 * at) {
// Really not enough prefs to meaningfully evaluate this user
continue;
}
FastIDSet relevantItemIDs = new FastIDSet(at);
// List some most-preferred items that would count as (most) "relevant" results
double theRelevanceThreshold = Double.isNaN(relevanceThreshold) ? computeThreshold(prefs) : relevanceThreshold;
prefs.sortByValueReversed();
for (int i = 0; i < size && relevantItemIDs.size() < at; i++) {
if (prefs.getValue(i) >= theRelevanceThreshold) {
relevantItemIDs.add(prefs.getItemID(i));
}
}
int numRelevantItems = relevantItemIDs.size();
if (numRelevantItems <= 0) {
continue;
}
FastByIDMap<PreferenceArray> trainingUsers = new FastByIDMap<PreferenceArray>(dataModel.getNumUsers());
LongPrimitiveIterator it2 = dataModel.getUserIDs();
while (it2.hasNext()) {
processOtherUser(userID, relevantItemIDs, trainingUsers, it2.nextLong(), dataModel);
}
DataModel trainingModel = dataModelBuilder == null ? new GenericDataModel(trainingUsers)
: dataModelBuilder.buildDataModel(trainingUsers);
Recommender recommender = recommenderBuilder.buildRecommender(trainingModel);
try {
trainingModel.getPreferencesFromUser(userID);
} catch (NoSuchUserException nsee) {
continue; // Oops we excluded all prefs for the user -- just move on
}
int intersectionSize = 0;
List<RecommendedItem> recommendedItems = recommender.recommend(userID, at, rescorer);
for (RecommendedItem recommendedItem : recommendedItems) {
if (relevantItemIDs.contains(recommendedItem.getItemID())) {
intersectionSize++;
}
}
int numRecommendedItems = recommendedItems.size();
// Precision
if (numRecommendedItems > 0) {
precision.addDatum((double) intersectionSize / (double) numRecommendedItems);
}
// Recall
recall.addDatum((double) intersectionSize / (double) numRelevantItems);
// Fall-out
if (numRelevantItems < size) {
fallOut.addDatum((double) (numRecommendedItems - intersectionSize)
/ (double) (numItems - numRelevantItems));
}
// nDCG
// In computing, assume relevant IDs have relevance 1 and others 0
double cumulativeGain = 0.0;
double idealizedGain = 0.0;
for (int i = 0; i < recommendedItems.size(); i++) {
RecommendedItem item = recommendedItems.get(i);
double discount = i == 0 ? 1.0 : 1.0 / log2(i + 1);
if (relevantItemIDs.contains(item.getItemID())) {
cumulativeGain += discount;
}
// otherwise we're multiplying discount by relevance 0 so it doesn't do anything
// Ideally results would be ordered with all relevant ones first, so this theoretical
// ideal list starts with number of relevant items equal to the total number of relevant items
if (i < relevantItemIDs.size()) {
idealizedGain += discount;
}
}
nDCG.addDatum(cumulativeGain / idealizedGain);
long end = System.currentTimeMillis();
log.info("Evaluated with user {} in {}ms", userID, end - start);
log.info("Precision/recall/fall-out/nDCG: {} / {} / {} / {}", new Object[] {
precision.getAverage(), recall.getAverage(), fallOut.getAverage(), nDCG.getAverage()
});
}
return new IRStatisticsImpl(precision.getAverage(), recall.getAverage(), fallOut.getAverage(), nDCG.getAverage());
}