package jsat.clustering.evaluation;
import java.util.List;
import jsat.DataSet;
import jsat.classifiers.ClassificationDataSet;
import jsat.classifiers.DataPoint;
import jsat.utils.DoubleList;
/**
* Normalized Mutual Information (NMI) is a measure to evaluate a cluster based
* on the true class labels for the data set. The NMI normally returns a value
* in [0, 1], where 0 indicates the clustering appears random, and 1 indicate
* the clusters perfectly match the class labels. To match the
* {@link ClusterEvaluation} interface, the value returned by evaluate will
* be 1.0-NMI .
*
* NOTE: Because the NMI needs to know the true class labels, only
* {@link #evaluate(int[], jsat.DataSet) } will work, since it provides the data
* set as an argument. The dataset given must be an instance of
* {@link ClassificationDataSet}
*
* @author Edward Raff
*/
public class NormalizedMutualInformation implements ClusterEvaluation
{
@Override
public double evaluate(int[] designations, DataSet dataSet)
{
if( !(dataSet instanceof ClassificationDataSet))
throw new RuntimeException("NMI can only be calcuate for classification data sets");
ClassificationDataSet cds = (ClassificationDataSet)dataSet;
double nmiNumer = 0.0;
double nmiC = 0.0;
double nmiK = 0.0;
DoubleList kPriors = new DoubleList();
for(int i= 0; i < cds.getSampleSize(); i++)
{
int ki = designations[i];
if(ki < 0)//outlier, not clustered
continue;
while(kPriors.size() <= ki)
kPriors.add(0.0);
kPriors.set(ki, kPriors.get(ki)+cds.getDataPoint(i).getWeight());
}
double N = 0.0;
for(int i = 0; i < kPriors.size(); i++)
N += kPriors.get(i);
for(int i = 0; i < kPriors.size(); i++)
{
kPriors.set(i, kPriors.get(i)/N);
double pKi = kPriors.get(i);
if(pKi > 0)
nmiK += - pKi*Math.log(pKi);
}
double[] cPriors = cds.getPriors();
double[][] ck = new double[cPriors.length][kPriors.size()];
for(int i = 0; i < cds.getSampleSize(); i++)
{
int ci = cds.getDataPointCategory(i);
int kj = designations[i];
if(kj < 0)//outlier, ignore
continue;
ck[ci][kj] += cds.getDataPoint(i).getWeight();
}
for(int i = 0; i < cPriors.length; i++)
{
double pCi = cPriors[i];
if(pCi <= 0.0)
continue;
double logPCi = Math.log(pCi);
for(int j = 0; j < kPriors.size(); j++)
{
double pKj = kPriors.get(j);
if(pKj <= 0.0)
continue;
double pCiKj = ck[i][j]/N;
if(pCiKj <= 0.0)
continue;
nmiNumer += pCiKj* (Math.log(pCiKj) - Math.log(pKj) - logPCi);
}
nmiC += -pCi*logPCi;
}
return 1.0-nmiNumer/((nmiC+nmiK)/2);
}
@Override
public double evaluate(List> dataSets)
{
throw new UnsupportedOperationException("NMI requires the true data set"
+ " labels, call evaluate(int[] designations, DataSet dataSet)"
+ " instead");
}
@Override
public NormalizedMutualInformation clone()
{
return new NormalizedMutualInformation();
}
}
Ads help maintain this website.