package jsat.lossfunctions;
/**
* The HuberLoss loss function for regression. The HuberLoss loss switches between
* {@link SquaredLoss} and {@link AbsoluteLoss} loss based on a threshold value.
*
* This function is only partially twice differentiable.
*
* @author Edward Raff
*/
public class HuberLoss implements LossR
{
private static final long serialVersionUID = -4463269746356262940L;
private double c;
/**
* Creates a new HuberLoss loss
*
* @param c the threshold to switch between the squared and logistic loss at
*/
public HuberLoss(double c)
{
this.c = c;
}
/**
* Creates a new HuberLoss loss thresholded at 1
*/
public HuberLoss()
{
this(1);
}
/**
* Computes the HuberLoss loss
*
* @param pred the predicted value
* @param y the true value
* @param c the threshold value
* @return the HuberLoss loss
*/
public static double loss(double pred, double y, double c)
{
final double x = y - pred;
if (Math.abs(x) <= c)
return x * x * 0.5;
else
return c * (Math.abs(x) - c / 2);
}
/**
* Computes the first derivative of the HuberLoss loss
*
* @param pred the predicted value
* @param y the true value
* @param c the threshold value
* @return the first derivative of the HuberLoss loss
*/
public static double deriv(double pred, double y, double c)
{
double x = pred-y;
if (Math.abs(x) <= c)
return x;
else
return c * Math.signum(x);
}
/**
* Computes the second derivative of the HuberLoss loss, which only exists for
* values < {@code c}
*
* @param pred the predicted value
* @param y the target value
* @param c the threshold value
* @return the second derivative of the HuberLoss loss
*/
public static double deriv2(double pred, double y, double c)
{
if (Math.abs(pred-y) < c)
return 1;
else
return 0;
}
public static double regress(double score)
{
return score;
}
@Override
public double getLoss(double pred, double y)
{
return loss(pred, y, c);
}
@Override
public double getDeriv(double pred, double y)
{
return deriv(pred, y, c);
}
@Override
public double getDeriv2(double pred, double y)
{
return deriv2(pred, y, c);
}
@Override
public double getConjugate(double b, double pred, double y)
{
//from "Regression tasks in machine learning via Fenchel duality"
if(-c < b && b < c)
return b*b*0.5+b*y;
return Double.POSITIVE_INFINITY;
}
@Override
public double getDeriv2Max()
{
return 1;
}
@Override
public HuberLoss clone()
{
return new HuberLoss(c);
}
@Override
public double getRegression(double score)
{
return score;
}
@Override
public double lipschitz()
{
//TODO: is this tighter?
//We also observe that |fc''(x)| ≤ 1/c. This implies that fc(x) is 1/c-Lipschitz continuous.
//Huber interpolated between Asbolute loss and Squared loss, with are both L-Lipschitz, so Huber is too
return 1;
}
}
Ads help maintain this website.