package jsat.lossfunctions;
/**
* The ε-insensitive loss for regression L(x, y) =
* max(0, |x-y|-ε) is the common loss function used for Support
* Vector Regression.
* When ε = 0, the loss becomes equivalent to the {@link AbsoluteLoss}.
*
* @author Edward Raff
*/
public class EpsilonInsensitiveLoss implements LossR
{
private static final long serialVersionUID = -8735274561429676350L;
/**
* Computes the ε-insensitive loss
*
* @param pred the predicted value
* @param y the true value
* @param eps the epsilon tolerance
* @return the ε-insensitive loss
*/
public static double loss(double pred, double y, double eps)
{
final double x = Math.abs(pred - y);
return Math.max(0, x-eps);
}
/**
* Computes the first derivative of the ε-insensitive loss
*
* @param pred the predicted value
* @param y the true value
* @param eps the epsilon tolerance
* @return the first derivative of the ε-insensitive loss
*/
public static double deriv(double pred, double y, double eps)
{
final double x = pred - y;
if(eps < Math.abs(x))
return Math.signum(x);
else
return 0;
}
private double eps;
/**
* Creates a new Epsilon Insensitive loss
* @param eps the epsilon tolerance on error
*/
public EpsilonInsensitiveLoss(double eps)
{
if(eps < 0 || Double.isNaN(eps) || Double.isInfinite(eps))
throw new IllegalArgumentException("Epsilon must be non-negative, not " + eps);
this.eps = eps;
}
/**
* Copy constructor
* @param toCopy the object to copy
*/
public EpsilonInsensitiveLoss(EpsilonInsensitiveLoss toCopy)
{
this.eps = toCopy.eps;
}
@Override
public double getLoss(double pred, double y)
{
return loss(pred, y, eps);
}
@Override
public double getDeriv(double pred, double y)
{
return deriv(pred, y, eps);
}
@Override
public double getDeriv2(double pred, double y)
{
return 0;
}
@Override
public double getConjugate(double b, double pred, double y)
{
//from "Regression tasks in machine learning via Fenchel duality"
if(Math.abs(b) < eps)
return b*y+eps*Math.abs(b);
return Double.POSITIVE_INFINITY;
}
@Override
public double getDeriv2Max()
{
return 0;
}
@Override
public EpsilonInsensitiveLoss clone()
{
return new EpsilonInsensitiveLoss(this);
}
@Override
public double getRegression(double score)
{
return score;
}
@Override
public double lipschitz()
{
//see "Are Loss Functions All the Same?"
return 1;
}
}
Ads help maintain this website.