private void calculateBuckets(
Set<? extends Person> persons,
DynamicDoubleArray sums,
DynamicIntArray counts,
String xAttrKey,
String yAttrKey) {
TIntDoubleHashMap sumBuckets = new TIntDoubleHashMap();
TIntIntHashMap countBuckets = new TIntIntHashMap();
for (Person person : persons) {
String xValStr = person.getAttribute(xAttrKey);
String yValStr = person.getAttribute(yAttrKey);
if (xValStr != null && yValStr != null) {
double xVal = Double.parseDouble(xValStr);
double yVal = Double.parseDouble(yValStr);
int bucketIdx = xDataDiscr.index(xVal);
sumBuckets.adjustOrPutValue(bucketIdx, yVal, yVal);
countBuckets.adjustOrPutValue(bucketIdx, 1, 1);
}
}
TIntDoubleIterator it = sumBuckets.iterator();
for (int i = 0; i < sumBuckets.size(); i++) {
it.advance();
int bucketIndex = it.key();
double sum = it.value();
int cnt = countBuckets.get(bucketIndex);
sums.set(bucketIndex, sum);
counts.set(bucketIndex, cnt);
}
}
@Override
public TIntDoubleHashMap getRating(
Document document, List<TokenSequence<SemanticEntity>> entities) {
DirectedGraph<Integer, RDFEdge> graph = document.getGraph();
TIntHashSet nodes = new TIntHashSet();
for (TokenSequence<SemanticEntity> ts : entities) {
nodes.add(ts.getValue().getSubjectIndex());
}
final TIntDoubleHashMap m = new TIntDoubleHashMap();
for (int node : nodes.toArray()) {
m.put(node, Math.min(graph.getOutEdges(node).size(), graph.getInEdges(node).size()));
}
return m;
}
/**
* For a specific sub-set of blocks (child nodes), find a 'base' subset of parents for which the
* block's logLikelihood is not -Infinity
*
* @param candidateParentsPerNode
* @param chosenArcsPerNode
* @param setOfBlocks
* @return
*/
protected double getOutOfMinusInfinity(
Int2ObjectOpenHashMap<IntOpenHashSet> candidateParentsPerNode,
Int2ObjectOpenHashMap<ObjectOpenHashSet<Arc>> chosenArcsPerNode,
IntOpenHashSet setOfBlocks,
TIntDoubleHashMap logLPerNode) {
double totalLogL = 0;
ProgressLogger pl = new ProgressLogger(LOGGER, ProgressLogger.TEN_SECONDS, "blocks");
pl.start("Begin initializing, to avoid zero likelihood, using set-cover heuristic");
pl.expectedUpdates = setOfBlocks.size();
int nArcs = 0;
for (int v : setOfBlocks) {
pl.update();
IntOpenHashSet vParents = candidateParentsPerNode.get(v);
Int2ObjectOpenHashMap<IntOpenHashSet> parentActions =
new Int2ObjectOpenHashMap<IntOpenHashSet>();
Int2ObjectOpenHashMap<IntArrayList> cPlusV = auxiliary.getCplusOnline(v);
Int2ObjectOpenHashMap<IntArrayList> cMinusV = auxiliary.getCminusOnline(v);
if (cPlusV != null) {
IntSet actions = cPlusV.keySet();
// Heuristic: first add the parents that participate in A+ for
// most actions
for (int action : actions) {
for (int u : cPlusV.get(action)) {
if (!parentActions.containsKey(u)) {
parentActions.put(u, new IntOpenHashSet());
}
parentActions.get(u).add(action);
}
}
}
KeepMaximum km = new KeepMaximum();
km.addAllKey2Listsize(parentActions);
IntOpenHashSet baseSetOfParents = new IntOpenHashSet();
double logL = Double.NEGATIVE_INFINITY;
while (logL == Double.NEGATIVE_INFINITY && (km.getMaximumKey() != -1)) {
int u = km.getMaximumKey();
if (baseSetOfParents.contains(u)) {
throw new IllegalStateException("Attempted to add twice the same parent");
}
baseSetOfParents.add(u);
logL = blockLogLikelihood(v, cPlusV, cMinusV, baseSetOfParents);
IntOpenHashSet uActions = parentActions.get(u);
for (int parent : vParents) {
parentActions.get(parent).removeAll(uActions);
}
vParents.remove(u);
parentActions.remove(u);
km.reset();
km.addAllKey2Listsize(parentActions);
}
// keep track of the likelihood
totalLogL += logL;
if (logLPerNode != null) {
logLPerNode.put(v, logL);
}
chosenArcsPerNode.put(v, new ObjectOpenHashSet<Arc>());
for (int u : baseSetOfParents) {
nArcs++;
chosenArcsPerNode.get(v).add(new Arc(u, v));
}
}
pl.stop("Done initialization. Added " + nArcs + " arcs, logLikelihood=" + totalLogL);
return totalLogL;
}
