hex

Class Distribution

java.lang.Object

water.Iced<Distribution>

hex.Distribution

All Implemented Interfaces:

java.io.Externalizable, java.io.Serializable, java.lang.Cloneable, Freezable<Distribution>

public abstract class Distribution
extends Iced<Distribution>

Distribution functions to be used by ML Algos

See Also:

Serialized Form

Field Summary

Fields

Modifier and Type	Field and Description
hex.genmodel.utils.DistributionFamily	distribution
double	huberDelta
hex.LinkFunction	linkFunction
double	quantileAlpha
double	tweediePower

Constructor Summary

Constructors

Constructor and Description
Distribution(hex.genmodel.utils.DistributionFamily family)
Distribution(hex.genmodel.utils.DistributionFamily family, hex.LinkFunction lf)
Distribution(Model.Parameters params)
Distribution(Model.Parameters params, hex.LinkFunction lf)

Method Summary

Modifier and Type	Method and Description
double	deviance(double w, double y, double f) Deviance of given distribution function at predicted value f
double	gammaDenom(double w, double y, double z, double f) Contribution to denominator for GBM's leaf node prediction
double	gammaNum(double w, double y, double z, double f) Contribution to numerator for GBM's leaf node prediction
double	initFDenom(double w, double o, double y) Contribution to denominator for initial value computation
double	initFNum(double w, double o, double y) Contribution to numerator for initial value computation
double	link(double f) Canonical link
double	linkInv(double f) Canonical link inverse
java.lang.String	linkInvString(java.lang.String f) String version of link inverse (for POJO scoring code generation)
double	negHalfGradient(double y, double f) (Negative half) Gradient of deviance function at predicted value f, for actual response y This assumes that the deviance(w,y,f) is w*deviance(y,f), so the gradient is w * d/df deviance(y,f)
void	setHuberDelta(double huberDelta) Setter of huber delta.

Methods inherited from class water.Iced

asBytes, clone, copyOver, frozenType, read, readExternal, readJSON, reloadFromBytes, toJsonString, write, writeExternal, writeJSON

Methods inherited from class java.lang.Object

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

Field Detail

tweediePower

public final double tweediePower

quantileAlpha

public final double quantileAlpha

huberDelta

public double huberDelta

linkFunction

public final hex.LinkFunction linkFunction

distribution

public final hex.genmodel.utils.DistributionFamily distribution

Constructor Detail

Distribution

public Distribution(hex.genmodel.utils.DistributionFamily family,
                    hex.LinkFunction lf)

Distribution

public Distribution(hex.genmodel.utils.DistributionFamily family)

Distribution

public Distribution(Model.Parameters params,
                    hex.LinkFunction lf)

Distribution

public Distribution(Model.Parameters params)

Method Detail

setHuberDelta

public void setHuberDelta(double huberDelta)

Setter of huber delta. Required for Huber aka M-regression.

Parameters:

huberDelta -

link

public double link(double f)

Canonical link

Parameters:

f - value in original space, to be transformed to link space

Returns:

link(f)

linkInv

public double linkInv(double f)

Canonical link inverse

Parameters:

f - value in link space, to be transformed back to original space

Returns:

linkInv(f)

linkInvString

public java.lang.String linkInvString(java.lang.String f)

String version of link inverse (for POJO scoring code generation)

Parameters:

f - value to be transformed by link inverse

Returns:

String that turns into compilable expression of linkInv(f)

deviance

public double deviance(double w,
                       double y,
                       double f)

Deviance of given distribution function at predicted value f

Parameters:

w - observation weight

y - (actual) response

f - (predicted) response in original response space (including offset)

Returns:

deviance

negHalfGradient

public double negHalfGradient(double y,
                              double f)

(Negative half) Gradient of deviance function at predicted value f, for actual response y This assumes that the deviance(w,y,f) is w*deviance(y,f), so the gradient is w * d/df deviance(y,f)

Parameters:

y - (actual) response

f - (predicted) response in link space (including offset)

Returns:

-1/2 * d/df deviance(w=1,y,f)

initFNum

public double initFNum(double w,
                       double o,
                       double y)

Contribution to numerator for initial value computation

Parameters:

w - weight

o - offset

y - response

Returns:

weighted contribution to numerator

initFDenom

public double initFDenom(double w,
                         double o,
                         double y)

Contribution to denominator for initial value computation

Parameters:

w - weight

o - offset

y - response

Returns:

weighted contribution to denominator

gammaNum

public double gammaNum(double w,
                       double y,
                       double z,
                       double f)

Contribution to numerator for GBM's leaf node prediction

Parameters:

w - weight

y - response

z - residual

f - predicted value (including offset)

Returns:

weighted contribution to numerator

gammaDenom

public double gammaDenom(double w,
                         double y,
                         double z,
                         double f)

Contribution to denominator for GBM's leaf node prediction

Parameters:

w - weight

y - response

z - residual

f - predicted value (including offset)

Returns:

weighted contribution to denominator