@Test
public void testBasicClustering() {
List<? extends WeightedVector> data = cubishTestData(1);
BallKMeans r = new BallKMeans(new BruteSearch(new EuclideanDistanceMeasure()), 6, 20);
r.cluster(data);
for (Centroid centroid : r) {
for (int i = 0; i < 10; i++) {
System.out.printf("%10.4f", centroid.get(i));
}
System.out.printf("\n");
}
}
@Test
public void testClustering() {
UpdatableSearcher searcher = new BruteSearch(new EuclideanDistanceMeasure());
BallKMeans clusterer = new BallKMeans(searcher, 1 << NUM_DIMENSIONS, NUM_ITERATIONS);
long startTime = System.currentTimeMillis();
clusterer.cluster(syntheticData.getFirst());
long endTime = System.currentTimeMillis();
assertEquals(
"Total weight not preserved",
totalWeight(syntheticData.getFirst()),
totalWeight(clusterer),
1e-9);
// Verify that each corner of the cube has a centroid very nearby.
// This is probably FALSE for large-dimensional spaces!
double maxWeight = 0;
for (Vector mean : syntheticData.getSecond()) {
WeightedThing<Vector> v = searcher.search(mean, 1).get(0);
maxWeight = Math.max(v.getWeight(), maxWeight);
}
assertTrue("Maximum weight too large " + maxWeight, maxWeight < 0.05);
double clusterTime = (endTime - startTime) / 1000.0;
System.out.printf(
"%s\n%.2f for clustering\n%.1f us per row\n\n",
searcher.getClass().getName(),
clusterTime,
clusterTime / syntheticData.getFirst().size() * 1e6);
// verify that the total weight of the centroids near each corner is correct
double[] cornerWeights = new double[1 << NUM_DIMENSIONS];
Searcher trueFinder = new BruteSearch(new EuclideanDistanceMeasure());
for (Vector trueCluster : syntheticData.getSecond()) {
trueFinder.add(trueCluster);
}
for (Centroid centroid : clusterer) {
WeightedThing<Vector> closest = trueFinder.search(centroid, 1).get(0);
cornerWeights[((Centroid) closest.getValue()).getIndex()] += centroid.getWeight();
}
int expectedNumPoints = NUM_DATA_POINTS / (1 << NUM_DIMENSIONS);
for (double v : cornerWeights) {
System.out.printf("%f ", v);
}
System.out.println();
for (double v : cornerWeights) {
assertEquals(expectedNumPoints, v, 0);
}
}
@Test
public void testInitialization() {
// start with super clusterable data
List<? extends WeightedVector> data = cubishTestData(0.01);
// just do initialization of ball k-means. This should drop a point into each of the clusters
BallKMeans r = new BallKMeans(new BruteSearch(new EuclideanDistanceMeasure()), 6, 20);
r.cluster(data);
// put the centroids into a matrix
Matrix x = new DenseMatrix(6, 5);
int row = 0;
for (Centroid c : r) {
x.viewRow(row).assign(c.viewPart(0, 5));
row++;
}
// verify that each column looks right. Should contain zeros except for a single 6.
final Vector columnNorms =
x.aggregateColumns(
new VectorFunction() {
@Override
public double apply(Vector f) {
// return the sum of three discrepancy measures
return Math.abs(f.minValue())
+ Math.abs(f.maxValue() - 6)
+ Math.abs(f.norm(1) - 6);
}
});
// verify all errors are nearly zero
assertEquals(0, columnNorms.norm(1) / columnNorms.size(), 0.1);
// verify that the centroids are a permutation of the original ones
SingularValueDecomposition svd = new SingularValueDecomposition(x);
Vector s = svd.getS().viewDiagonal().assign(Functions.div(6));
assertEquals(5, s.getLengthSquared(), 0.05);
assertEquals(5, s.norm(1), 0.05);
}
