/** {@inheritDoc} */
public void processSpace(Properties properties) {
SparseMatrix cleanedMatrix = (SparseMatrix) transform.transform(cooccurrenceMatrix);
for (String term : basis.keySet()) {
int index = basis.getDimension(term);
SparseDoubleVector sdv = cleanedMatrix.getRowVector(index);
double score = 0;
for (int i : sdv.getNonZeroIndices()) score += sdv.get(i);
wordScores.put(term, score);
}
}
/**
* Returns an array containing the row indices of the neighbors of the impost node and the row
* index of the impost node itself.
*/
private static int[] getImpostNeighbors(SparseMatrix sm, int rowIndex) {
int[] impost1edges = sm.getRowVector(rowIndex).getNonZeroIndices();
int[] neighbors = Arrays.copyOf(impost1edges, impost1edges.length + 1);
neighbors[neighbors.length - 1] = rowIndex;
return neighbors;
}
/**
* {@inheritDoc}
*
* @throws IllegalArgumentException if {@code matrix} is not square, or is not an instance of
* {@link SparseMatrix}
*/
public Assignments cluster(Matrix matrix, Properties props) {
if (matrix.rows() != matrix.columns())
throw new IllegalArgumentException(
"Input matrix is not square. "
+ "Matrix is expected to be a square matrix whose values (i,j) "
+ "denote an edge from row i to row j");
if (!(matrix instanceof SparseMatrix)) {
throw new IllegalArgumentException("Input matrix must be a " + "sparse matrix.");
}
SparseMatrix sm = (SparseMatrix) matrix;
String inMemProp = props.getProperty(KEEP_SIMILARITY_MATRIX_IN_MEMORY_PROPERTY);
boolean keepSimMatrixInMem = (inMemProp != null) ? Boolean.parseBoolean(inMemProp) : true;
// IMPLEMENTATION NOTE: Ahn et al. used single-linkage HAC, which can be
// efficiently implemented in O(n^2) time as a special case of HAC.
// However, we currently don't optimize for this special case and
// instead use our HAC class. Because of the complexity of the edge
// similarity function, we build our own similarity matrix and then pass
// it in, rather than passing in the edge matrix directly.
final int rows = sm.rows();
numRows = rows;
LOGGER.fine("Generating link similarity matrix for " + rows + " nodes");
// Rather than create an O(row^3) matrix for representing the edges,
// compress the edge matrix by getting a mapping for each edge to a row
// in the new matrix.
final List<Edge> edgeList = new ArrayList<Edge>();
this.edgeList = edgeList;
for (int r = 0; r < rows; ++r) {
SparseDoubleVector row = sm.getRowVector(r);
int[] edges = row.getNonZeroIndices();
for (int col : edges) {
// Always add edges from the upper triangular
if (r > col) edgeList.add(new Edge(r, col));
// Otherwise, we only add the edge from the lower triangular if
// it wasn't present in the upper. This avoids counting
// duplicate edges.
else if (r < col && sm.get(col, r) == 0) edgeList.add(new Edge(r, col));
}
}
final int numEdges = edgeList.size();
LOGGER.fine("Number of edges to cluster: " + numEdges);
Matrix edgeSimMatrix = getEdgeSimMatrix(edgeList, sm, keepSimMatrixInMem);
LOGGER.fine("Computing single linkage link clustering");
final List<Merge> mergeOrder =
new HierarchicalAgglomerativeClustering()
.buildDendrogram(edgeSimMatrix, ClusterLinkage.SINGLE_LINKAGE);
this.mergeOrder = mergeOrder;
LOGGER.fine("Calculating partition densitities");
// Set up a concurrent map that each thread will update once it has
// calculated the densitites of each of its partitions. This map is
// only written to once per thread.
final ConcurrentNavigableMap<Double, Integer> partitionDensities =
new ConcurrentSkipListMap<Double, Integer>();
// Register a task group for calculating all of the partition
// densitities
Object key = WORK_QUEUE.registerTaskGroup(mergeOrder.size());
for (int p = 0; p < mergeOrder.size(); ++p) {
final int part = p;
WORK_QUEUE.add(
key,
new Runnable() {
public void run() {
// Get the merges for this particular partitioning of
// the links
List<Merge> mergeSteps = mergeOrder.subList(0, part);
// Convert the merges to a specific cluster labeling
MultiMap<Integer, Integer> clusterToElements =
convertMergesToAssignments(mergeSteps, numEdges);
// Based on the link partitioning, calculate the
// partition density for each cluster
double partitionDensitySum = 0d;
for (Integer cluster : clusterToElements.keySet()) {
Set<Integer> linkPartition = clusterToElements.get(cluster);
int numLinks = linkPartition.size();
BitSet nodesInPartition = new BitSet(rows);
for (Integer linkIndex : linkPartition) {
Edge link = edgeList.get(linkIndex);
nodesInPartition.set(link.from);
nodesInPartition.set(link.to);
}
int numNodes = nodesInPartition.cardinality();
// This reflects the density of this particular
// cluster
double partitionDensity =
(numLinks - (numNodes - 1d))
/ (((numNodes * (numNodes - 1d)) / 2d) - (numLinks - 1));
partitionDensitySum += partitionDensity;
}
// Compute the density for the total partitioning
// solution
double partitionDensity = (2d / numEdges) * partitionDensitySum;
LOGGER.log(
Level.FINER,
"Partition solution {0} had " + "density {1}",
new Object[] {part, partitionDensity});
// Update the thread-shared partition density map with
// this task's calculation
partitionDensities.put(partitionDensity, part);
}
});
}
// Wait for all the partition densities to be calculated
WORK_QUEUE.await(key);
Map.Entry<Double, Integer> densest = partitionDensities.lastEntry();
LOGGER.fine(
"Partition " + densest.getValue() + " had the highest density: " + densest.getKey());
int partitionWithMaxDensity = densest.getValue();
// Select the solution with the highest partition density and assign
// nodes accordingly
MultiMap<Integer, Integer> bestEdgeAssignment =
convertMergesToAssignments(mergeOrder.subList(0, partitionWithMaxDensity), numEdges);
List<Set<Integer>> nodeClusters = new ArrayList<Set<Integer>>(rows);
for (int i = 0; i < rows; ++i) nodeClusters.add(new HashSet<Integer>());
// Ignore the original partition labeling, and use our own cluster
// labeling to ensure that the IDs are contiguous.
int clusterId = 0;
// For each of the partitions, add the partion's cluster ID to all the
// nodes that are connected by one of the partition's edges
for (Integer cluster : bestEdgeAssignment.keySet()) {
Set<Integer> edgePartition = bestEdgeAssignment.get(cluster);
for (Integer edgeId : edgePartition) {
Edge e = edgeList.get(edgeId);
nodeClusters.get(e.from).add(clusterId);
nodeClusters.get(e.to).add(clusterId);
}
// Update the cluster id
clusterId++;
}
int numClusters = 0;
Assignment[] nodeAssignments = new Assignment[rows];
for (int i = 0; i < nodeAssignments.length; ++i) {
nodeAssignments[i] = new SoftAssignment(nodeClusters.get(i));
}
return new Assignments(numClusters, nodeAssignments, matrix);
}