Skip to content

GH-16678 GLM: Control variables - Multinomial, Ordinal #16690

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
maurever wants to merge 12 commits into
base: rel-3.46.0
Choose a base branch
from
Open

GH-16678 GLM: Control variables - Multinomial, Ordinal #16690

Show file tree
Hide file tree
Changes from 4 commits
Commits
Show all changes
12 commits
Select commit Hold shift + click to select a range
48c0744
Implement multinomial case for control variables.
maurever Nov 13, 2025
abe358b
Fix junit test for multinomial control variables
maurever Dec 1, 2025
d1b2c0f
Add pyunit, runit, fix bugs.
maurever Dec 2, 2025
8388b6b
Implement ordinal distribution, add tests
maurever Dec 8, 2025
48634c9
Update h2o-algos/src/main/java/hex/glm/GLM.java
maurever Dec 8, 2025
a49cabb
Update h2o-algos/src/main/java/hex/glm/GLMModel.java
maurever Dec 8, 2025
644638f
Update h2o-r/tests/testdir_algos/glm/runit_GLM_control_variables_ordi…
maurever Dec 8, 2025
e4b310b
Update h2o-r/tests/testdir_algos/glm/runit_GLM_control_variables_mult…
maurever Dec 8, 2025
b78bf38
Update h2o-r/tests/testdir_algos/glm/runit_GLM_control_variables_ordi…
maurever Dec 8, 2025
70d9250
Update h2o-algos/src/test/java/hex/glm/GLMControlVariablesTest.java
maurever Dec 8, 2025
005c0aa
Improve tests
maurever Dec 9, 2025
5cd1a8e
Improve tests
maurever Dec 10, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
3 changes: 2 additions & 1 deletion h2o-algos/src/main/java/hex/api/MakeGLMModelHandler.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,6 +15,7 @@
import water.api.schemas3.KeyV3;
import water.fvec.*;
import water.fvec.Vec.VectorGroup;
import water.util.ArrayUtils;

import java.util.Arrays;
import java.util.Map;
Expand Down Expand Up @@ -77,7 +78,7 @@ public GLMModelV3 make_unrestricted_model(int version, MakeUnrestrictedGLMModelV
DataInfo dinfo = model.dinfo();
dinfo.setPredictorTransform(TransformType.NONE);
m._output = new GLMOutput(model.dinfo(), model._output._names, model._output._column_types, model._output._domains,
model._output.coefficientNames(), model._output.beta(), model._output._binomial, model._output._multinomial,
model._output.coefficientNames(), model._output._multinomial || model._output._ordinal? ArrayUtils.flattenArray(model._output.get_global_beta_multinomial()) : model._output.beta(), model._output._binomial, model._output._multinomial,
model._output._ordinal, null);
ModelMetrics mt = model._output._training_metrics_unrestricted_model;
ModelMetrics mv = model._output._validation_metrics_unrestricted_model;
Expand Down
34 changes: 24 additions & 10 deletions h2o-algos/src/main/java/hex/glm/GLM.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,6 +2,7 @@

import Jama.Matrix;
import hex.*;
import hex.genmodel.utils.DistributionFamily;
import hex.glm.GLMModel.GLMOutput;
import hex.glm.GLMModel.GLMParameters.Family;
import hex.glm.GLMModel.GLMParameters.Link;
Expand Down Expand Up @@ -3400,17 +3401,23 @@ private void scorePostProcessingControlVal(Frame train, long t1) {
_model.addScoringInfo(_parms, nclasses(), t2, _state._iter); // add to scoringInfo for early stopping

if (_parms._generate_scoring_history) { // update scoring history with deviance train and valid if available
double likelihoodContrVal;
double[] betaContrVal = _model._output.getControlValBeta(_state.expandBeta(_state.beta()).clone());
GLMResDevTask task = new GLMResDevTask(_job._key, _dinfo, _parms, betaContrVal).doAll(_dinfo._adaptedFrame);
double objectiveControlVal = _state.objective(betaContrVal, task._likelihood);

if(_model._output.isMultinomialClassifier()){
GLMResDevTaskMultinomial task = new GLMResDevTaskMultinomial(_job._key, _dinfo, betaContrVal, nclasses()).doAll(_dinfo._adaptedFrame);
likelihoodContrVal = task._likelihood;
} else {
GLMResDevTask task = new GLMResDevTask(_job._key, _dinfo, _parms, betaContrVal).doAll(_dinfo._adaptedFrame);
likelihoodContrVal = task._likelihood;
}
double objectiveControlVal = _state.objective(betaContrVal, likelihoodContrVal);
if ((mtrain != null) && (_valid != null)) {
_scoringHistory.addIterationScore(true, true, _state._iter, task._likelihood,
objectiveControlVal, _state.deviance(task._likelihood), ((GLMMetrics) _model._output._validation_metrics).residual_deviance(),
_scoringHistory.addIterationScore(true, true, _state._iter, likelihoodContrVal,
objectiveControlVal, _state.deviance(likelihoodContrVal), ((GLMMetrics) _model._output._validation_metrics).residual_deviance(),
mtrain._nobs, _model._output._validation_metrics._nobs, _state.lambda(), _state.alpha());
} else { // only doing training deviance
_scoringHistory.addIterationScore(true, false, _state._iter, task._likelihood,
objectiveControlVal, _state.deviance(task._likelihood), Double.NaN, mtrain._nobs, 1, _state.lambda(),
_scoringHistory.addIterationScore(true, false, _state._iter, likelihoodContrVal,
objectiveControlVal, _state.deviance(likelihoodContrVal), Double.NaN, mtrain._nobs, 1, _state.lambda(),
Comment thread
maurever marked this conversation as resolved.
_state.alpha());
}
_job.update(_workPerIteration, _state.toString());
Expand Down Expand Up @@ -4036,9 +4043,16 @@ protected void updateProgress(boolean canScore) {
if (_model._parms._control_variables != null){
_scoringHistoryUnrestrictedModel.addIterationScore(_state._iter, _state.likelihood(), _state.objective());
double[] betaContrVal = _model._output.getControlValBeta(_state.expandBeta(_state.beta()).clone());
GLMResDevTask task = new GLMResDevTask(_job._key,_dinfo,_parms, betaContrVal).doAll(_state._dinfo._adaptedFrame);
double objectiveControlVal = _state.objective(betaContrVal, task._likelihood);
_scoringHistory.addIterationScore(_state._iter, task._likelihood, objectiveControlVal);
double likelihoosContrVal;
if(_model._output.isMultinomialClassifier()){
GLMResDevTaskMultinomial task = new GLMResDevTaskMultinomial(_job._key, _dinfo, betaContrVal, nclasses()).doAll(_state._dinfo._adaptedFrame);
likelihoosContrVal = task._likelihood;
} else {
GLMResDevTask task = new GLMResDevTask(_job._key,_dinfo,_parms, betaContrVal).doAll(_state._dinfo._adaptedFrame);
likelihoosContrVal = task._likelihood;
}
double objectiveControlVal = _state.objective(betaContrVal, likelihoosContrVal);
_scoringHistory.addIterationScore(_state._iter, likelihoosContrVal, objectiveControlVal);
Comment thread
maurever marked this conversation as resolved.
Outdated
} else {
_scoringHistory.addIterationScore(_state._iter, _state.likelihood(), _state.objective());
}
Expand Down
42 changes: 35 additions & 7 deletions h2o-algos/src/main/java/hex/glm/GLMModel.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1646,6 +1646,28 @@ public double[] getControlValBeta(double[] beta){
}
return beta;
}

public double[][] getControlValBetaMultinomial(double[][] beta) {
if (_control_values_idxs_in_adapted_frame == null) {
mapControlVariables();
}
assert _control_values_idxs_in_adapted_frame != null;
for (int featureIdx : _control_values_idxs_in_adapted_frame) {
if (featureIdx < _dinfo._catOffsets.length - 1 && _column_types[featureIdx].equals("Enum")) {
Copy link
Copy Markdown
Contributor

@tomasfryda tomasfryda Jan 2, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

&& _column_types[featureIdx].equals("Enum")

Is this part really necessary or is it just to be really really sure?

for (int i = _dinfo._catOffsets[featureIdx]; i < _dinfo._catOffsets[featureIdx + 1]; i++) {
for (int c = 0; c < beta.length; ++c) {
beta[c][i] = 0;
}
}
} else {
for (int c = 0; c < beta.length; ++c) {
featureIdx += _dinfo._numOffsets[0] - _dinfo._catOffsets.length + 1;
beta[c][featureIdx] = 0;
Comment thread
maurever marked this conversation as resolved.
Outdated
}
}
}
return beta;
}

public double[] stdErr() {
return calculateStdErrFromZValues(_zvalues, _global_beta);
Expand Down Expand Up @@ -1797,7 +1819,6 @@ public GLMOutput(DataInfo dinfo, String[] column_names, String[] column_types, S
mapControlVariables();
}


public GLMOutput() {
_isSupervised = true;
_nclasses = -1;
Expand Down Expand Up @@ -2020,9 +2041,13 @@ public VarImp calculateVarimp(boolean contrVal) {
float[] magnitudesSort = new float[len]; // stored sorted coefficient magnitudes
String[] namesSort = new String[len];

if (contrVal)
calculateVarimpBase(magnitudes, indices, getControlValBeta(getNormBeta()));
else if (_nclasses > 2)
if (contrVal) {
if(_nclasses > 2) {
calculateVarimpMultinomial(magnitudes, indices, getControlValBetaMultinomial(getNormBetaMultinomial()));
} else {
calculateVarimpBase(magnitudes, indices, getControlValBeta(getNormBeta()));
}
} else if (_nclasses > 2)
calculateVarimpMultinomial(magnitudes, indices, getNormBetaMultinomial());
else
calculateVarimpBase(magnitudes, indices, getNormBeta());
Expand Down Expand Up @@ -2147,9 +2172,13 @@ else if (_output.bestSubmodel().alpha_value == 1)
int icptInd = bm[0].length-1;
if (_parms._family == Family.ordinal) // only need one eta for all classes
classInd -= 1; // last class all zeros
double[][] bmcv = bm.clone();
if(_useControlVariables){
bmcv = _output.getControlValBetaMultinomial(bm);
}
for (int c = 0; c < classInd; ++c) {
double e = bm[c][icptInd]; // grab the intercept, replace the bm[0].length-1
double [] b = bm[c];
double e = bmcv[c][icptInd]; // grab the intercept, replace the bm[0].length-1
double [] b = bmcv[c];
for(int i = 0; i < _output._dinfo._cats; ++i) {
int l = _output._dinfo.getCategoricalId(i, data[i]);
if (l >= 0) e += b[l];
Expand Down Expand Up @@ -2203,7 +2232,6 @@ else if (_output.bestSubmodel().alpha_value == 1)
double[] bcv = b.clone();
if (this._useControlVariables)
bcv = _output.getControlValBeta(bcv); // make beta connected to control variables zero

for (int i = 0; i < _output._dinfo._cats && !Double.isNaN(eta); ++i) {
int l = _output._dinfo.getCategoricalId(i, data[i]);
if (l >= 0) eta += bcv[l];
Expand Down
9 changes: 6 additions & 3 deletions h2o-algos/src/main/java/hex/glm/GLMScore.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -47,7 +47,11 @@ public GLMScore(Job j, GLMModel m, DataInfo dinfo, String[] domain, boolean comp
if(_m._parms._family == GLMModel.GLMParameters.Family.multinomial ||
_m._parms._family == GLMModel.GLMParameters.Family.ordinal){
_beta = null;
_beta_multinomial = m._output._global_beta_multinomial;
if(m._useControlVariables){
_beta_multinomial = m._output.getControlValBetaMultinomial(m._output._global_beta_multinomial);
} else {
_beta_multinomial = m._output._global_beta_multinomial;
}
_dinfo = dinfo;
} else {
double[] beta = m.beta();
Expand All @@ -67,7 +71,6 @@ public GLMScore(Job j, GLMModel m, DataInfo dinfo, String[] domain, boolean comp
beta2[l] = beta[beta.length - 1];
beta = beta2;
}

if (m._useControlVariables) {
double[] betaContVar = m.beta().clone();
betaContVar = m._output.getControlValBeta(betaContVar);
Expand Down Expand Up @@ -116,7 +119,7 @@ public GLMScore(Job j, GLMModel m, DataInfo dinfo, String[] domain, boolean comp
preds[_nclasses] = 1-previousCDF;
preds[0] = ArrayUtils.maxIndex(preds)-1;
} else if (_m._parms._family == GLMModel.GLMParameters.Family.multinomial) {
double[] eta = _eta;
double[] eta = _eta.clone();
final double[][] bm = _beta_multinomial;
double sumExp = 0;
double maxRow = 0;
Expand Down
Loading