public DistanceDependentCRPGibbsOperator(
Parameter links,
Parameter assignments,
Parameter chiParameter,
NPAntigenicLikelihood Likelihood,
double weight) {
this.links = links;
this.assignments = assignments;
this.modelLikelihood = Likelihood;
this.chiParameter = chiParameter;
this.depMatrix = Likelihood.getLogDepMatrix();
for (int i = 0; i < links.getDimension(); i++) {
links.setParameterValue(i, i);
}
setWeight(weight);
// double[][] x=modelLikelihood.getData();
// modelLikelihood.printInformtion(x[0][0]);
this.m = new double[2];
m[0] = modelLikelihood.priorMean.getParameterValue(0);
m[1] = modelLikelihood.priorMean.getParameterValue(1);
this.v0 = 2;
this.k0 =
modelLikelihood.priorPrec.getParameterValue(0)
/ modelLikelihood.clusterPrec.getParameterValue(0);
this.T0Inv = new double[2][2];
T0Inv[0][0] = v0 / modelLikelihood.clusterPrec.getParameterValue(0);
T0Inv[1][1] = v0 / modelLikelihood.clusterPrec.getParameterValue(0);
T0Inv[1][0] = 0.0;
T0Inv[0][1] = 0.0;
this.logDetT0 = -Math.log(T0Inv[0][0] * T0Inv[1][1]);
}
/** change the parameter and return the hastings ratio. */
public final double doOperation() {
int index = MathUtils.nextInt(links.getDimension());
int oldGroup = (int) assignments.getParameterValue(index);
/*
* Set index customer link to index and all connected to it to a new assignment (min value empty)
*/
int minEmp = minEmpty(modelLikelihood.getLogLikelihoodsVector());
links.setParameterValue(index, index);
int[] visited = connected(index, links);
int ii = 0;
while (visited[ii] != 0) {
assignments.setParameterValue(visited[ii] - 1, minEmp);
ii++;
}
/*
* Adjust likvector for group separated
*/
modelLikelihood.setLogLikelihoodsVector(oldGroup, getLogLikGroup(oldGroup));
modelLikelihood.setLogLikelihoodsVector(minEmp, getLogLikGroup(minEmp));
int maxFull = maxFull(modelLikelihood.getLogLikelihoodsVector());
double[] liks = modelLikelihood.getLogLikelihoodsVector();
/*
* computing likelihoods of joint groups
*/
double[] crossedLiks = new double[maxFull + 1];
for (int ll = 0; ll < maxFull + 1; ll++) {
if (ll != minEmp) {
crossedLiks[ll] = getLogLik2Group(ll, minEmp);
}
}
/*
* Add logPrior
*/
double[] logP = new double[links.getDimension()];
for (int jj = 0; jj < links.getDimension(); jj++) {
logP[jj] += depMatrix[index][jj];
int n = (int) assignments.getParameterValue(jj);
if (n != minEmp) {
logP[jj] += crossedLiks[n] - liks[n] - liks[minEmp];
}
}
logP[index] = Math.log(chiParameter.getParameterValue(0));
/*
* possibilidade de mandar p zero as probs muito pequenas
*/
/*
* Gibbs sampling
*/
this.rescale(logP); // Improve numerical stability
this.exp(logP); // Transform back to probability-scale
int k = MathUtils.randomChoicePDF(logP);
links.setParameterValue(index, k);
int newGroup = (int) assignments.getParameterValue(k);
ii = 0;
while (visited[ii] != 0) {
assignments.setParameterValue(visited[ii] - 1, newGroup);
ii++;
}
/*
* updating conditional likelihood vector
*/
modelLikelihood.setLogLikelihoodsVector(newGroup, getLogLikGroup(newGroup));
if (newGroup != minEmp) {
modelLikelihood.setLogLikelihoodsVector(minEmp, 0);
}
sampleMeans(maxFull);
return 0.0;
}
public void sampleMeans(int maxFull) {
double[][] means = new double[maxFull + 2][2];
// sample mean vector for each cluster
for (int i = 0; i < maxFull + 1; i++) {
// Find all elements in cluster
int ngroup = 0;
for (int ii = 0; ii < assignments.getDimension(); ii++) {
if ((int) assignments.getParameterValue(ii) == i) {
ngroup++;
}
}
if (ngroup != 0) {
double[][] group = new double[ngroup][2];
double[] groupMean = new double[2];
int count = 0;
for (int ii = 0; ii < assignments.getDimension(); ii++) {
if ((int) assignments.getParameterValue(ii) == i) {
group[count][0] = modelLikelihood.getData()[ii][0];
group[count][1] = modelLikelihood.getData()[ii][1];
groupMean[0] += group[count][0];
groupMean[1] += group[count][1];
count += 1;
}
}
groupMean[0] /= ngroup;
groupMean[1] /= ngroup;
double kn = k0 + ngroup;
double vn = v0 + ngroup;
double[][] sumdif = new double[2][2];
for (int jj = 0; jj < ngroup; jj++) {
sumdif[0][0] += (group[jj][0] - groupMean[0]) * (group[jj][0] - groupMean[0]);
sumdif[0][1] += (group[jj][0] - groupMean[0]) * (group[jj][1] - groupMean[1]);
sumdif[1][0] += (group[jj][0] - groupMean[0]) * (group[jj][1] - groupMean[1]);
sumdif[1][1] += (group[jj][1] - groupMean[1]) * (group[jj][1] - groupMean[1]);
}
double[][] TnInv = new double[2][2];
TnInv[0][0] =
T0Inv[0][0]
+ ngroup * (k0 / kn) * (groupMean[0] - m[0]) * (groupMean[0] - m[0])
+ sumdif[0][0];
TnInv[0][1] =
T0Inv[0][1]
+ ngroup * (k0 / kn) * (groupMean[1] - m[1]) * (groupMean[0] - m[0])
+ sumdif[0][1];
TnInv[1][0] =
T0Inv[1][0]
+ ngroup * (k0 / kn) * (groupMean[0] - m[0]) * (groupMean[1] - m[1])
+ sumdif[1][0];
TnInv[1][1] =
T0Inv[1][1]
+ ngroup * (k0 / kn) * (groupMean[1] - m[1]) * (groupMean[1] - m[1])
+ sumdif[1][1];
Matrix Tn = new SymmetricMatrix(TnInv).inverse();
double[] posteriorMean = new double[2];
// compute posterior mean
posteriorMean[0] = (k0 * m[0] + ngroup * groupMean[0]) / (k0 + ngroup);
posteriorMean[1] = (k0 * m[1] + ngroup * groupMean[1]) / (k0 + ngroup);
// compute posterior Precision
double[][] posteriorPrecision =
new WishartDistribution(vn, Tn.toComponents()).nextWishart();
posteriorPrecision[0][0] *= kn;
posteriorPrecision[1][0] *= kn;
posteriorPrecision[0][1] *= kn;
posteriorPrecision[1][1] *= kn;
double[] sample =
new MultivariateNormalDistribution(posteriorMean, posteriorPrecision)
.nextMultivariateNormal();
means[i][0] = sample[0];
means[i][1] = sample[1];
}
}
// Fill in cluster means for each observation
for (int j = 0; j < assignments.getDimension(); j++) {
double[] group = new double[2];
group[0] = means[(int) assignments.getParameterValue(j)][0];
group[1] = means[(int) assignments.getParameterValue(j)][1];
modelLikelihood.setMeans(j, group);
}
}
public double getLogLik2Group(int group1, int group2) {
double L = 0.0;
int ngroup1 = 0;
for (int i = 0; i < assignments.getDimension(); i++) {
if ((int) assignments.getParameterValue(i) == group1) {
ngroup1++;
}
}
int ngroup2 = 0;
for (int i = 0; i < assignments.getDimension(); i++) {
if ((int) assignments.getParameterValue(i) == group2) {
ngroup2++;
}
}
int ngroup = (ngroup1 + ngroup2);
if (ngroup != 0) {
double[][] group = new double[ngroup][2];
double mean[] = new double[2];
int count = 0;
for (int i = 0; i < assignments.getDimension(); i++) {
if ((int) assignments.getParameterValue(i) == group1) {
group[count][0] = modelLikelihood.getData()[i][0];
group[count][1] = modelLikelihood.getData()[i][1];
mean[0] += group[count][0];
mean[1] += group[count][1];
count += 1;
}
}
for (int i = 0; i < assignments.getDimension(); i++) {
if ((int) assignments.getParameterValue(i) == group2) {
group[count][0] = modelLikelihood.getData()[i][0];
group[count][1] = modelLikelihood.getData()[i][1];
mean[0] += group[count][0];
mean[1] += group[count][1];
count += 1;
}
}
mean[0] /= ngroup;
mean[1] /= ngroup;
double kn = k0 + ngroup;
double vn = v0 + ngroup;
double[][] sumdif = new double[2][2];
for (int i = 0; i < ngroup; i++) {
sumdif[0][0] += (group[i][0] - mean[0]) * (group[i][0] - mean[0]);
sumdif[0][1] += (group[i][0] - mean[0]) * (group[i][1] - mean[1]);
sumdif[1][0] += (group[i][0] - mean[0]) * (group[i][1] - mean[1]);
sumdif[1][1] += (group[i][1] - mean[1]) * (group[i][1] - mean[1]);
}
double[][] TnInv = new double[2][2];
TnInv[0][0] =
T0Inv[0][0] + ngroup * (k0 / kn) * (mean[0] - m[0]) * (mean[0] - m[0]) + sumdif[0][0];
TnInv[0][1] =
T0Inv[0][1] + ngroup * (k0 / kn) * (mean[1] - m[1]) * (mean[0] - m[0]) + sumdif[0][1];
TnInv[1][0] =
T0Inv[1][0] + ngroup * (k0 / kn) * (mean[0] - m[0]) * (mean[1] - m[1]) + sumdif[1][0];
TnInv[1][1] =
T0Inv[1][1] + ngroup * (k0 / kn) * (mean[1] - m[1]) * (mean[1] - m[1]) + sumdif[1][1];
double logDetTn = -Math.log(TnInv[0][0] * TnInv[1][1] - TnInv[0][1] * TnInv[1][0]);
L += -(ngroup) * Math.log(Math.PI);
L += Math.log(k0) - Math.log(kn);
L += (vn / 2) * logDetTn - (v0 / 2) * logDetT0;
L += GammaFunction.lnGamma(vn / 2) + GammaFunction.lnGamma((vn / 2) - 0.5);
L += -GammaFunction.lnGamma(v0 / 2) - GammaFunction.lnGamma((v0 / 2) - 0.5);
}
return L;
}