hex.glm

Class DispersionUtils

java.lang.Object

hex.glm.DispersionUtils

public class DispersionUtils
extends java.lang.Object

Constructor Summary

Constructors

Constructor and Description
DispersionUtils()

Method Summary

Modifier and Type	Method and Description
static double	dispersionLS(DispersionTask.ComputeMaxSumSeriesTsk computeTsk, TweedieMLDispersionOnly tDispersion, GLMModel.GLMParameters parms)
static double	estimateGammaMLSE(GLMTask.ComputeGammaMLSETsk mlCT, double seOld, double[] beta, GLMModel.GLMParameters parms, ComputationState state, water.Job job, GLMModel model) Estimate dispersion factor using maximum likelihood.
static double	estimateNegBinomialDispersionFisherScoring(GLMModel.GLMParameters parms, GLMModel model, double[] beta, DataInfo dinfo)
static double	estimateNegBinomialDispersionMomentMethod(GLMModel model, double[] beta, DataInfo dinfo, water.fvec.Vec weights, water.fvec.Vec response, water.fvec.Vec mu)
static double	estimateTweedieDispersionOnly(GLMModel.GLMParameters parms, GLMModel model, water.Job job, double[] beta, DataInfo dinfo) This method estimates the tweedie dispersion parameter.
static double[]	makeZeros(double[] sourceCoeffs, double[] targetCoeffs)

Methods inherited from class java.lang.Object

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

Constructor Detail

DispersionUtils

public DispersionUtils()

Method Detail

estimateGammaMLSE

public static double estimateGammaMLSE(GLMTask.ComputeGammaMLSETsk mlCT,
                                       double seOld,
                                       double[] beta,
                                       GLMModel.GLMParameters parms,
                                       ComputationState state,
                                       water.Job job,
                                       GLMModel model)

Estimate dispersion factor using maximum likelihood. I followed section IV of the doc in https://github.com/h2oai/h2o-3/issues/7013.

estimateTweedieDispersionOnly

public static double estimateTweedieDispersionOnly(GLMModel.GLMParameters parms,
                                                   GLMModel model,
                                                   water.Job job,
                                                   double[] beta,
                                                   DataInfo dinfo)

This method estimates the tweedie dispersion parameter. It will use Newton's update if the new update will increase the loglikelihood. Otherwise, the dispersion will be updated as dispersionNew = dispersionCurr + learningRate * update. In addition, line search is used to increase the magnitude of the update when the update magnitude is too small (< 1e-3). Every 10th iteration it checks if the optimization doesn't diverge. If it looks like it diverged, it uses a different likelihood estimation that should be more accurate (combination of Series and Fourier inversion method) but without gradients. For this reason it will use a Golden section search which doesn't require gradients and has a linear convergence. For details, please see sections IV.I, IV.II, and IV.III in document here:

estimateNegBinomialDispersionMomentMethod

public static double estimateNegBinomialDispersionMomentMethod(GLMModel model,
                                                               double[] beta,
                                                               DataInfo dinfo,
                                                               water.fvec.Vec weights,
                                                               water.fvec.Vec response,
                                                               water.fvec.Vec mu)

estimateNegBinomialDispersionFisherScoring

public static double estimateNegBinomialDispersionFisherScoring(GLMModel.GLMParameters parms,
                                                                GLMModel model,
                                                                double[] beta,
                                                                DataInfo dinfo)

dispersionLS

public static double dispersionLS(DispersionTask.ComputeMaxSumSeriesTsk computeTsk,
                                  TweedieMLDispersionOnly tDispersion,
                                  GLMModel.GLMParameters parms)

makeZeros

public static double[] makeZeros(double[] sourceCoeffs,
                                 double[] targetCoeffs)