@Override
protected KMeans fit() {
synchronized (fitLock) {
if (null != labels) // already fit
return this;
final LogTimer timer = new LogTimer();
final double[][] X = data.getData();
final int n = data.getColumnDimension();
final double nan = Double.NaN;
// Corner case: K = 1 or all singular values
if (1 == k) {
labelFromSingularK(X);
fitSummary.add(new Object[] {iter, converged, tss, tss, nan, timer.wallTime()});
sayBye(timer);
return this;
}
// Nearest centroid model to predict labels
NearestCentroid model = null;
EntryPair<int[], double[]> label_dist;
// Keep track of TSS (sum of barycentric distances)
double last_wss_sum = Double.POSITIVE_INFINITY, wss_sum = 0;
ArrayList<double[]> new_centroids;
for (iter = 0; iter < maxIter; iter++) {
// Get labels for nearest centroids
try {
model =
new NearestCentroid(
CentroidUtils.centroidsToMatrix(centroids, false),
VecUtils.arange(k),
new NearestCentroidParameters()
.setSeed(getSeed())
.setMetric(getSeparabilityMetric())
.setVerbose(false))
.fit();
} catch (NaNException NaN) {
/*
* If they metric used produces lots of infs or -infs, it
* makes it hard if not impossible to effectively segment the
* input space. Thus, the centroid assignment portion below can
* yield a zero count (denominator) for one or more of the centroids
* which makes the entire row NaN. We should tell the user to
* try a different metric, if that's the case.
*
error(new IllegalClusterStateException(dist_metric.getName()+" produced an entirely " +
"infinite distance matrix, making it difficult to segment the input space. Try a different " +
"metric."));
*/
this.k = 1;
warn(
"(dis)similarity metric ("
+ dist_metric
+ ") cannot partition space without propagating Infs. Returning one cluster");
labelFromSingularK(X);
fitSummary.add(new Object[] {iter, converged, tss, tss, nan, timer.wallTime()});
sayBye(timer);
return this;
}
label_dist = model.predict(X);
// unpack the EntryPair
labels = label_dist.getKey();
new_centroids = new ArrayList<>(k);
int label;
wss = new double[k];
int[] centroid_counts = new int[k];
double[] centroid;
double[][] new_centroid_arrays = new double[k][n];
for (int i = 0; i < m; i++) {
label = labels[i];
centroid = centroids.get(label);
// increment count for this centroid
double this_cost = 0;
centroid_counts[label]++;
for (int j = 0; j < centroid.length; j++) {
double diff = X[i][j] - centroid[j];
this_cost += (diff * diff);
// Add the the centroid sums
new_centroid_arrays[label][j] += X[i][j];
}
// add this cost to the WSS
wss[label] += this_cost;
}
// one pass of K for some consolidation
wss_sum = 0;
for (int i = 0; i < k; i++) {
wss_sum += wss[i];
for (int j = 0; j < n; j++) // meanify
new_centroid_arrays[i][j] /= (double) centroid_counts[i];
new_centroids.add(new_centroid_arrays[i]);
}
// update the BSS
bss = tss - wss_sum;
// Assign new centroids
double diff = last_wss_sum - wss_sum;
last_wss_sum = wss_sum;
// Check for convergence and add summary:
converged = FastMath.abs(diff) < tolerance; // first iter will be inf
fitSummary.add(
new Object[] {
converged ? iter++ : iter, converged, tss, wss_sum, bss, timer.wallTime()
});
if (converged) {
break;
} else {
// otherwise, reassign centroids
centroids = new_centroids;
}
} // end iterations
// Reorder the labels, centroids and wss indices
reorderLabelsAndCentroids();
if (!converged) warn("algorithm did not converge");
// wrap things up, create summary..
sayBye(timer);
return this;
}
}