jsat.lossfunctions

Class HuberLoss

java.lang.Object

jsat.lossfunctions.HuberLoss

All Implemented Interfaces:

java.io.Serializable, LossFunc, LossR

public class HuberLoss
extends java.lang.Object
implements LossR

The HuberLoss loss function for regression. The HuberLoss loss switches between SquaredLoss and AbsoluteLoss loss based on a threshold value.
This function is only partially twice differentiable.

See Also:

Serialized Form

Constructor Summary

Constructors

Constructor and Description
HuberLoss() Creates a new HuberLoss loss thresholded at 1
HuberLoss(double c) Creates a new HuberLoss loss

Method Summary

Modifier and Type	Method and Description
HuberLoss	clone()
static double	deriv(double pred, double y, double c) Computes the first derivative of the HuberLoss loss
static double	deriv2(double pred, double y, double c) Computes the second derivative of the HuberLoss loss, which only exists for values < c
double	getConjugate(double b, double pred, double y) Computes the result of the conjugate function of this loss.
double	getDeriv(double pred, double y) Computes the first derivative of the getLoss function.
double	getDeriv2(double pred, double y) Computes the second derivative of the getLoss function.
double	getDeriv2Max() Returns an upper bound on the maximum value of the second derivative.
double	getLoss(double pred, double y) Computes the getLoss for a regression problem.
double	getRegression(double score) Given the score value of a data point, this returns the correct numeric result.
double	lipschitz() If this loss is L-Lipschitz (1/L Lipschitz smooth), this method will return the value of L.
static double	loss(double pred, double y, double c) Computes the HuberLoss loss
static double	regress(double score)

Methods inherited from class java.lang.Object

equals, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Constructor Detail

HuberLoss

public HuberLoss(double c)

Creates a new HuberLoss loss

Parameters:

c - the threshold to switch between the squared and logistic loss at

HuberLoss

public HuberLoss()

Creates a new HuberLoss loss thresholded at 1

Method Detail

loss

public static double loss(double pred,
                          double y,
                          double c)

Computes the HuberLoss loss

Parameters:

pred - the predicted value

y - the true value

c - the threshold value

Returns:

the HuberLoss loss

deriv

public static double deriv(double pred,
                           double y,
                           double c)

Computes the first derivative of the HuberLoss loss

Parameters:

pred - the predicted value

y - the true value

c - the threshold value

Returns:

the first derivative of the HuberLoss loss

deriv2

public static double deriv2(double pred,
                            double y,
                            double c)

Computes the second derivative of the HuberLoss loss, which only exists for values < c

Parameters:

pred - the predicted value

y - the target value

c - the threshold value

Returns:

the second derivative of the HuberLoss loss

regress

public static double regress(double score)

getLoss

public double getLoss(double pred,
                      double y)

Description copied from interface: LossR

Computes the getLoss for a regression problem.

Specified by:

getLoss in interface LossFunc

Specified by:

getLoss in interface LossR

Parameters:

pred - the predicted value in (-Infinity, Infinity)

y - the true target value in (-Infinity, Infinity)

Returns:

the getLoss in [0, Inf)

getDeriv

public double getDeriv(double pred,
                       double y)

Description copied from interface: LossR

Computes the first derivative of the getLoss function.

Specified by:

getDeriv in interface LossFunc

Specified by:

getDeriv in interface LossR

Parameters:

pred - the predicted value in (-Infinity, Infinity)

y - the true target value in (-Infinity, Infinity)

Returns:

the first derivative of the getLoss

getDeriv2

public double getDeriv2(double pred,
                        double y)

Description copied from interface: LossR

Computes the second derivative of the getLoss function.

Specified by:

getDeriv2 in interface LossFunc

Specified by:

getDeriv2 in interface LossR

Parameters:

pred - the predicted value in (-Infinity, Infinity)

y - the true target value in (-Infinity, Infinity)

Returns:

the second derivative of the getLoss function

getConjugate

public double getConjugate(double b,
                           double pred,
                           double y)

Description copied from interface: LossFunc

Computes the result of the conjugate function of this loss. This function is generally optional, and should return Double.NaN if not properly implemented. Many optimization algorithms do require a working implementation though.

Specified by:

getConjugate in interface LossFunc

Parameters:

b - the primary input to the function

pred - the predicted score in (-Infinity, Infinity)

y - the true class label in {-1, 1}

Returns:

the result of the conjugate function of this loss

getDeriv2Max

public double getDeriv2Max()

Description copied from interface: LossFunc

Returns an upper bound on the maximum value of the second derivative. If the second derivative does not exist, Double.NaN is a valid result. It is also possible for 0 and Double.POSITIVE_INFINITY to be valid results, and must be checked for.

Specified by:

getDeriv2Max in interface LossFunc

Returns:

the max value of LossFunc.getDeriv2(double, double)

clone

public HuberLoss clone()

Specified by:

clone in interface LossFunc

Specified by:

clone in interface LossR

Overrides:

clone in class java.lang.Object

getRegression

public double getRegression(double score)

Description copied from interface: LossR

Given the score value of a data point, this returns the correct numeric result. For most regression problems this simply returns the score value.

Specified by:

getRegression in interface LossR

Parameters:

score - the score for a data point

Returns:

the correct numeric regression value for this loss function

lipschitz

public double lipschitz()

Description copied from interface: LossFunc

If this loss is L-Lipschitz (1/L Lipschitz smooth), this method will return the value of L. If it is not L-Lipschitz, a value of 0 will be returned.

Specified by:

lipschitz in interface LossFunc

Returns:

the L-Lipschitz constant, or 0 if this loss is not L-Lipschitz;