water.udf

Interface CDistributionFunc

public interface CDistributionFunc

Custom Distribution Function Interface to customize loss and prediction calculation in GBM algorithm The function has four parts: - link: link function transforms the probability of response variable to a continuous scale that is unbounded - init: computes numerator and denominator of the initial value. - gradient: computes (Negative half) Gradient of deviance function at predicted value for actual response - gamma: computes numerator and denominator of terminal node estimate

Method Summary

Modifier and Type	Method and Description
double[]	gamma(double w, double y, double z, double f) Contribution for GBM's leaf node prediction (numerator and denominator).
double	gradient(double y, double f) (Negative half) Gradient of deviance function at predicted value f, for actual response y.
double	gradient(double y, double f, int l) (Negative half) Gradient of deviance function at predicted value f, for actual response y.
double[]	init(double w, double o, double y) Contribution for initial value computation (numerator and denominator).
java.lang.String	link() Type of Link function.

Method Detail

link

java.lang.String link()

Type of Link function.

Returns:

name of link function. Possible functions: log, logit, identity, inverse, ologit, ologlog, oprobit

init

double[] init(double w,
              double o,
              double y)

Contribution for initial value computation (numerator and denominator).

Parameters:

w - weight

o - offset

y - response

Returns:

[weighted contribution to init numerator, weighted contribution to init denominator]

gradient

double gradient(double y,
                double f)

(Negative half) Gradient of deviance function at predicted value f, for actual response y. Important for customization of a loss function.

Parameters:

y - (actual) response

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

Returns:

gradient

gradient

double gradient(double y,
                double f,
                int l)

(Negative half) Gradient of deviance function at predicted value f, for actual response y. Important for customization of a loss function.

Parameters:

y - (actual) response

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

l - (class label) label of a class (converted lexicographically from original labels to 0-number of class - 1)

Returns:

gradient

gamma

double[] gamma(double w,
               double y,
               double z,
               double f)

Contribution for GBM's leaf node prediction (numerator and denominator). Important for customization of a loss function.

Parameters:

w - weight

y - response

z - residual

f - predicted value (including offset)

Returns:

[weighted contribution to gamma numerator, weighted contribution to gamma denominator]