What's New in H2O Machine Learning: Christmas 2018

There were two releases shortly after one another. First, on December 21st, there was a minor (fix) release 3.22.0.3 immediately followed by a more major release (but still on 3.22 branch) codename Xu, named after mathematician Jinchao Xu, whose work is focused on deep neural networks besides many other fields of research.

Of course, the new 3.22.1.1 release with codename Xu contains all the fixes present in the previous fix release, 3.22.0.3. The following points are highlights of the most impactful changes. For a complete list of changes, fixes, and improvements, please read the recent changes section.

Support for CDH 6.x

H₂O now declares support for Cloudera distributions based on Hadoop 3. This includes releases by Cloudera CDH 6.0, CDH 6.1. Support for releases by Hortonworks HDP 3.0 and HDP 3.1are going to be added shortly in one of the releases to follow.

Partial Dependence Plots Can Be Exported

The API to plot partial dependence graphs has been enhanced with the ability to directly export (save to disk) the plots. For both R and Python API, a new parameter named save_to_file has been added. The parameter accepts a string representing a path on a filesystem.

Python

The existing method partial_plot callable on any Model able has been modified by adding the new save_to_file parameter. All the plots are saved into one image file.

model.partial_plot(data = data,cols = ['AGE', 'RACE', 'DCAPS'],server = True, plot = True, save_to_file = "/home/username/pdp.png")

R

H₂O provides function h2o.partialPlot in R to create the plots. A new optional parameter named save_to has been introduced for this function. In R, different plots are traditionally represented by a separate pointer to a separate image. To honor this contract in H₂O R API, the plots are not saved into one file, but one file is created per each feature.

h2o.partialPlot(object = model, data = data, save_to = "/home/username/pdp")

If the model mentioned in the showcase above has been trained using a set of three features: [Age, RACE, DCAPS], then H2O is going to save three files onto the filesystem:

- /home/username/pdp_AGE.png
- /home/username/pdp_RACE.png
- /home/username/pdp_DCAPS.png

That is the reason why the string containing the filesystem path does not include png suffix. It is added automatically. If the suffix is added accidentally, it is stripped by H₂O automatically.

Monotonicity Constraints for GBM

H₂O users now have the ability to affect GBM splits by stating monotonic relationships of a feature to the predicted variable. Any subset of features used during model training phase might be restricted with monotonicity constraints.

R

A new argument named monotone_constraints has been added to h2o.gbm(..) function. This argument is optional and accepts a list of key:value. The key is the name of the feature and value is one of {-1,0,1}, where -1 represents decreasing constraint, 0 represents no constraint and 1 represents increasing constraint.

features <- ("Origin", "Dest","Distance")
features.constraints <- c(1, 0, -1)
monotonicity.constraints <- setNames(features.constraints, features)
gbm.model <- h2o.gbm(y = "IsDepDelayed",
                         training_frame = training_frame,
                         monotone_constraints = monotonicity.constraints,
                         validation_frame = validation_frame)

Python

A new argument named monotone_constraints has been added to the H2OXGBoostEstimator’s constructor. This argument is optional and accepts a list of key:value. The key is the name of the feature and value is one of {-1,0,1}, where -1 represents decreasing constraint, 0 represents no constraint and 1 represents increasing constraint.

feature_names = ['MedInc', 'AveOccup', 'HouseAge']
monotone_constraints = {"MedInc": 1, "AveOccup": -1, "HouseAge": 1}
xgb_mono = H2OXGBoostEstimator(monotone_constraints = monotone_constraints)
xgb_mono.train(x = feature_names, y = "target", training_frame = train, validation_frame = test)

Full Python demo is available on H₂O GitHub.

AutoML Performance Improvement

In version 3.22.0.3, the performance of all models in the AutoML run has been improved. AutoML previously automatically partitioned the training set, setting aside 10% of the data to be used for an early stopping. This dataset was not being used, so now AutoML simply uses the full training dataset to train the models, leading to better model performance. All details, including the actual fixes done, are to be found in PUBDEV-6079 JIRA.

Custom Metrics for Early Stopping

Specification for custom stopping metrics is now supported for GBM, DRF, and GLM algorithms.

There are two new stopping criteria:

• custom — for custom metric functions where “less is better”, it is expected that the lower bound is 0
• custom_increasing — for custom metric functions where “more is better”

Python

Custom stopping function

 def custom_stopping_metric_function():
     return h2o.upload_custom_metric(CustomMaeFunc, func_name = "mae", func_file = "mm_mae.py")


model_actual = H2OGradientBoostingEstimator(model_id="prostate", ntrees = 10, max_depth = 5,
                                                 score_each_iteration = True,
                                                 custom_metric_func = custom_stopping_metric_function(),
                                                 stopping_metric = "custom",
                                                 stopping_tolerance = 0.1,
                                                 stopping_rounds = 3)
     model_actual.train(y = "AGE", x = ftrain.names, training_frame = ftrain, validation_frame = fvalid)

Exporting Checkpoints in AutoML

Throughout the AutoML experiments and Grid searches, resulting models can now be checkpointed. By specifying export_checkpoints_dir value, which is a string pointing to a directory int he filesystem, checkpoints are saved automatically as new models are created.

AutoML Python

     model = H2OAutoML(project_name = "ExampleProject", stopping_rounds = 3, export_checkpoints_dir = "/home/username/example/checkpoints")
     model.train(y = "CAPSULE", training_frame = training_frame)

Grid Search Python

    air_grid = H2OGridSearch(H2OGradientBoostingEstimator, hyper_params = hyper_parameters, search_criteria = search_crit)
    air_grid.train(x = ["Origin", "Distance"], y = "IsDepDelayed", training_frame = training_frame, export_checkpoints_dir = checkpoints_dir)

AutoML R

    model <- h2o.automl(y = y,
                         training_frame = train,
                         project_name = "AutoMLTest", 
                         export_checkpoints_dir = "/home/username/example/checkpoints")

Grid search R

grid = h2o.grid("glm", grid_id = "glm_grid_cars_test", x = predictors, y = "economy", training_frame = train,
                            family = "gaussian", hyper_params = hyper_params, export_checkpoints_dir = "checkpoints_dir")