API-related changes

H2O-3 does its best to keep backwards compatibility between major versions, but sometimes breaking changes are needed in order to improve code quality and to address issues. This section provides a list of current breaking changes between specific releases.

From 3.32.0.1

Modules

The deprecated h2o-scala module has been removed.

Target Encoding

The Target Encoder API has been clarified and its consistency across clients has been improved. The following parameters are now deprecated in all clients and officially replaced by their new alternative:

• k \(\to\) inflection_point

• f \(\to\) smoothing

• noise_level \(\to\) noise

• use_blending (R only) \(\to\) blending

Legacy client code using the deprecated parameters should expect a deprecation warning when using them. You are strongly encouraged to update your code to use the new naming.

transform parameter updates

In an objective of performance optimization on the backend, and of simplification of the API, the transform method used to apply target encoding was modified as follows:

• The R h2o.transform function (accepting a target encoder model as the first argument) and the Python H2OTargetEncoderEstimator.transform methods are now fully compatible: they accept the same parameters and work consistently.

• The parameters data_leakage_handling, seed are now ignored on those methods: by default, transform will use the corresponding values defined when building the TargetEncoder model.

• The other regularization parameters on these transform methods (e.g. noise, blending, inflection_point, smoothing), always default to the value defined on the TargetEncoder model.

• A new as_training parameter has been introduced to simplify and enforce a correct usage of target encoding:

  • When transforming a training dataset, you should use (R) h2o.transform(te_model, train_dataset, as_training=TRUE) or (Python) te_model.transform(train_dataset, as_training=True).

  • When transforming any other dataset (validation, test, …), you can just use (R) h2o.transform(te_model, train_dataset) or (Python) te_model.transform(train_dataset).

  • Legacy code using for example h2o.transform(te_model, train_dataset, data_leakage_handling="KFold") will now be translated internally to h2o.transform(te_model, train_dataset, as_training=TRUE).

Finally the following APIs (deprecated since 3.28) have been fully removed:

• Python: h2o.targetencoder module.

• R: h2o.target_encode_fit and h2o.target_encode_transform functions.

Parameters

The max_hit_ratio_k parameter has been removed.

From 3.30.1.2

The max_hit_ratio_k parameter is deprecated in version 3.30.1.2 and will be completely removed in the next major version, 3.32.0.1.

From 3.30.1.1

The deprecated h2o-scala module has been removed.

From 3.30.0.5

The h2o-scala module is deprecated in version 3.30.0.5 and will be completely removed in the next major version, 3.30.1.1.

From 3.30.0.4

The following options are no longer supported by native XGBoost and have been removed.

• min_sum_hessian_in_leaf

• min_data_in_leaf

From 3.28 or below to 3.30

Java API

The hex.grid.HyperSpaceWalker and hex.grid.HyperspaceWalker.HyperSpaceIterator interfaces have been simplified. Users implementing those interfaces directly, for example to create a custom grid search exploration algorithm, may want to look at the default implementations in h2o-core/src/main/java/hex/grid/HyperSpaceWalker.java if they are facing any issue when compiling against the new interfaces.

From 3.26 or below to 3.28

Java API

The following classes were moved:

Until 3.26	From 3.28
ai.h2o.automl.EventLog	ai.h2o.automl.events.EventLog
ai.h2o.automl.EventLogEntry	ai.h2o.automl.events.EventLogEntry
ai.h2o.automl.Leaderboard	ai.h2o.automl.leaderboard.Leaderboard

From 3.22 or below to 3.24

Java API

The following classes were moved and/or renamed:

Until 3.22	From 3.24
hex.StackedEnsembleModel	hex.ensemble.StackedEnsembleModel
hex.StackedEnsembleModel.MetalearnerAlgorithm	hex.ensemble.Metalearner.Algorithm
ai.h2o.automl.AutoML.algo	ai.h2o.automl.Algo

Some internal methods of StackedEnsemble and StackedEnsembleModel are no longer public, but this should not impact you.