package jsat.clustering.evaluation;
import java.util.List;
import jsat.DataSet;
import jsat.classifiers.ClassificationDataSet;
import jsat.classifiers.DataPoint;
import static java.lang.Math.*;
/**
* Adjusted Rand Index (ARI) is a measure to evaluate a cluster based
* on the true class labels for the data set. The ARI normally returns a value
* in [-1, 1], where 0 indicates the clustering appears random, and 1 indicate
* the clusters perfectly match the class labels, and negative values indicate a
* clustering that is worse than random. To match the
* {@link ClusterEvaluation} interface, the value returned by evaluate will
* be 1.0-Adjusted Rand Index so the best value becomes 0.0 and the worse value
* becomes 2.0.
*
* NOTE: Because the ARI 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 AdjustedRandIndex 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;
int clusters = 0;//how many clusters are there?
for(int clusterID : designations)
clusters = Math.max(clusterID+1, clusters);
double[] truthSums = new double[cds.getClassSize()];
double[] clusterSums = new double[clusters];
double[][] table = new double[clusterSums.length][truthSums.length];
double n = 0.0;
for(int i = 0; i < designations.length; i++)
{
int cluster = designations[i];
if(cluster < 0)
continue;//noisy point
int label = cds.getDataPointCategory(i);
double weight = cds.getDataPoint(i).getWeight();
table[cluster][label] += weight;
truthSums[label] += weight;
clusterSums[cluster] += weight;
n += weight;
}
/*
* Adjusted Rand Index involves many (n choose 2) = 1/2 (n-1) n
*/
double sumAllTable = 0.0;
double addCTerm = 0.0, addLTerm = 0.0;//clustering and label
for(int i = 0; i < table.length; i++)
{
double a_i = clusterSums[i];
addCTerm += a_i*(a_i-1)/2;
for(int j = 0; j < table[i].length; j++)
{
if(i == 0)
{
double b_j = truthSums[j];
addLTerm += b_j*(b_j-1)/2;
}
double n_ij = table[i][j];
double n_ij_c2 = n_ij*(n_ij-1)/2;
sumAllTable += n_ij_c2;
}
}
double longMultTerm = exp(log(addCTerm)+log(addLTerm)-(log(n)+log(n-1)-log(2)));//numericaly more stable verison
return 1.0-(sumAllTable-longMultTerm)/(addCTerm/2+addLTerm/2-longMultTerm);
}
@Override
public double evaluate(List> dataSets)
{
throw new UnsupportedOperationException("Adjusted Rand Index requires the true data set"
+ " labels, call evaluate(int[] designations, DataSet dataSet)"
+ " instead");
}
@Override
public ClusterEvaluation clone()
{
return new AdjustedRandIndex();
}
}
Ads help maintain this website.