/**
* split the neighbor hood in two groups based on 2 k-means
*
* @param neighborhood
* @return
*/
private Pair<List<Gene>, List<Gene>> twoMeanClusterSplit(List<Gene> neighborhood) {
final int n = neighborhood.size();
final int maxit = desc.getMaxit();
final double eps = desc.getEps();
int a_start = r.nextInt(n);
int b_start = r.nextInt(n);
Gene a_center = new Gene(1, -1, Arrays.copyOf(neighborhood.get(a_start).data, samples));
Gene b_center = new Gene(1, -1, Arrays.copyOf(neighborhood.get(b_start).data, samples));
float[] a_center_pong = new float[samples];
Arrays.fill(a_center_pong, Float.NaN);
float[] b_center_pong = new float[samples];
Arrays.fill(b_center_pong, Float.NaN);
float[] tmp;
BitSet partOf_a = new BitSet(n);
double d_old = 0;
for (int i = 0; i < maxit; ++i) {
int j = 0;
int changed = 0;
double d_new = 0;
for (Gene gene : neighborhood) {
final double a_distance = distance(a_center, gene);
final double b_distance = distance(b_center, gene);
final boolean in_a = a_distance < b_distance;
if (partOf_a.get(j) != in_a) {
changed++;
partOf_a.set(j, in_a);
}
d_new += in_a ? a_distance : b_distance;
tmp = in_a ? a_center_pong : b_center_pong;
// shift new center
for (int k = 0; k < samples; ++k) {
if (!gene.isNaN(k)) {
if (Float.isNaN(tmp[k])) tmp[k] = gene.get(k);
else tmp[k] += gene.get(k);
}
}
j++;
}
if (changed == 0 || d_new == 0) break;
final double ratio = Math.abs(d_new - d_old) / d_old;
if (i > 0 && ratio < eps) break;
d_old = d_new;
int a_n = partOf_a.cardinality();
int b_n = n - a_n;
if (a_n == 0 || b_n == 0) {
// FIXME
}
updateCenter(a_center, a_center_pong, a_n);
updateCenter(b_center, b_center_pong, b_n);
}
return split(neighborhood, partOf_a);
}
private double distance(Gene target, Gene neighbor) {
double acc = 0;
int n = 0;
for (int sample = 0; sample < samples; ++sample) {
if (target.isNaN(sample) || neighbor.isNaN(sample)) // skip missing
continue;
double dx = target.get(sample) - neighbor.get(sample);
acc += dx * dx;
n++;
}
if (n > 0) {
return acc / n; // FIXME according to the fortran code, this is not the eucledian distance
// return Math.sqrt(acc);
}
return Double.POSITIVE_INFINITY;
}
private Sample computeSample(final int sample) {
int nans = 0;
double sum = 0;
int n = 0;
for (Gene gene : genes) {
double v = gene.get(sample);
if (isNaN(v)) nans++;
else {
sum += v;
n++;
}
}
return new Sample(sum / n, nans);
}
public double get(int sample) {
return gene.get(sample);
}
