package jsat.classifiers.evaluation;
import java.util.Arrays;
import jsat.classifiers.CategoricalData;
import jsat.classifiers.CategoricalResults;
/**
* This computes the multi-class Log Loss
* - 1/N Σ∀ i ∈ N log(pi, y)
*
*
* Where N is the number of data points and pi, y is
* the estimated probability of the true class label. The lower the loss score,
* the better.
*
* When pi, y = 0 the log loss is uninformatively forced to
* ∞, even if all other data points are perfectly correct. To avoid this a
* small nudge factor is added.
*
* @author Edward Raff
*/
public class LogLoss implements ClassificationScore
{
private static final long serialVersionUID = 3123851772991293430L;
private double loss;
private double weightSum;
private double nudge;
/**
* Creates a new Log Loss evaluation score
*/
public LogLoss()
{
this(1e-15);
}
/**
* Creates a new Log Loss evaluation score
* @param nudge the nudge value to avoid zero probabilities, must be non
* negative and less than 0.1
*/
public LogLoss(double nudge)
{
if(nudge < 0 || nudge >= 0.1)
throw new IllegalArgumentException("nudge must be a small non-negative value in [0, 0.1) not " + nudge);
this.nudge = nudge;
}
public LogLoss(LogLoss toClone)
{
this.loss = toClone.loss;
this.weightSum = toClone.weightSum;
this.nudge = toClone.nudge;
}
@Override
public void addResult(CategoricalResults prediction, int trueLabel, double weight)
{
loss += weight * Math.log(Math.max(prediction.getProb(trueLabel), nudge));
weightSum += weight;
}
@Override
public void addResults(ClassificationScore other)
{
LogLoss otherObj = (LogLoss) other;
this.loss += otherObj.loss;
this.weightSum += otherObj.weightSum;
}
@Override
public void prepare(CategoricalData toPredict)
{
loss = 0;
weightSum = 0;
}
@Override
public double getScore()
{
return -loss/weightSum;
}
@Override
public boolean equals(Object obj)
{
if(this.getClass().isAssignableFrom(obj.getClass()) && obj.getClass().isAssignableFrom(this.getClass()))
{
return ((LogLoss)obj).nudge == this.nudge;
}
return false;
}
@Override
public int hashCode()
{
return Arrays.hashCode(new double[]{nudge});
}
@Override
public boolean lowerIsBetter()
{
return true;
}
@Override
public LogLoss clone()
{
return new LogLoss(this);
}
@Override
public String getName()
{
return "Log Loss";
}
}
Ads help maintain this website.