Skip to main content
H2OAutoML automates the supervised machine learning model training process. It trains multiple models — including GBM, XGBoost, GLM, DRF, XRT, DeepLearning, and Stacked Ensembles — cross-validates them by default, and ranks them on a leaderboard. 
from h2o.automl import H2OAutoML

​
Constructor

H2OAutoML(
    nfolds=-1,
    balance_classes=False,
    class_sampling_factors=None,
    max_after_balance_size=5.0,
    max_runtime_secs=None,
    max_runtime_secs_per_model=None,
    max_models=None,
    distribution="AUTO",
    stopping_metric="AUTO",
    stopping_tolerance=None,
    stopping_rounds=3,
    seed=None,
    project_name=None,
    exclude_algos=None,
    include_algos=None,
    exploitation_ratio=-1,
    modeling_plan=None,
    preprocessing=None,
    monotone_constraints=None,
    keep_cross_validation_predictions=False,
    keep_cross_validation_models=False,
    keep_cross_validation_fold_assignment=False,
    sort_metric="AUTO",
    custom_metric_func=None,
    export_checkpoints_dir=None,
    verbosity="warn"
)

​
Stopping criteria

​
max_runtime_secs
integer
Maximum total time (in seconds) for the AutoML run. When both max_runtime_secs and max_models are specified, AutoML stops when either limit is reached. If neither is specified, defaults to 3600 seconds (1 hour).
​
max_runtime_secs_per_model
integer
default:"0"
Maximum time in seconds dedicated to each individual model. 0 disables the per-model limit.
​
max_models
integer
Maximum number of models to build, excluding Stacked Ensemble models. Set this parameter to ensure reproducibility: all models will be trained until convergence and none will be constrained by a time budget.
​
stopping_metric
string
default:"AUTO"
Metric used for early stopping during the AutoML run. "AUTO" resolves to "logloss" for classification and "deviance" for regression. Other options: "mse", "rmse", "mae", "rmsle", "auc", "aucpr", "misclassification", "mean_per_class_error", "r2".
​
stopping_tolerance
float
Relative tolerance for the metric-based stopping criterion. Defaults to 0.001 for datasets with at least 1 million rows; otherwise computed from dataset size.
​
stopping_rounds
integer
default:"3"
Stop training new models when the stopping metric has not improved for this many consecutive models. 0 disables this check.

​
Cross-validation

​
nfolds
integer
default:"-1"
Number of folds for k-fold cross-validation. -1 lets AutoML decide (uses 5-fold CV or a blending frame depending on dataset size). 0 disables CV. Minimum value when setting explicitly is 2.
​
keep_cross_validation_predictions
boolean
default:"False"
Retain cross-validation predictions. Required when continuing an existing AutoML project with repeated train() calls.
​
keep_cross_validation_models
boolean
default:"False"
Retain cross-validation sub-models. Keeping them consumes more cluster memory.

​
Algorithm selection

​
exclude_algos
string[]
Algorithms to skip. Available values: "DRF", "GLM", "XGBoost", "GBM", "DeepLearning", "StackedEnsemble". Cannot be combined with include_algos.
exclude_algos=["DeepLearning", "StackedEnsemble"]
​
include_algos
string[]
Restrict AutoML to only these algorithms. Cannot be combined with exclude_algos.
include_algos=["GBM", "XGBoost"]

​
Class balancing

​
balance_classes
boolean
default:"False"
Oversample minority classes to balance the class distribution. Only applicable for classification.
​
class_sampling_factors
float[]
Desired over/under-sampling ratios per class (in lexicographic order). Requires balance_classes=True. Auto-computed if not specified.
​
max_after_balance_size
float
default:"5.0"
Maximum relative size of the training dataset after class balancing. Majority classes are undersampled if the oversampled size exceeds this limit.

​
Reproducibility and project management

​
seed
integer
Random seed. AutoML guarantees reproducibility only when max_models or early stopping is used, because max_runtime_secs is resource-dependent.
​
project_name
string
Name for this AutoML project. Auto-generated from the training frame ID if not specified. Reuse the same name to continue training additional models on an existing project.

​
Leaderboard

​
sort_metric
string
default:"AUTO"
Metric used to sort the leaderboard. "AUTO" resolves to "auc" (binomial), "mean_per_class_error" (multinomial), or "deviance" (regression).

​
Other

​
distribution
string | dict
default:"AUTO"
Distribution family used by supporting algorithms. Options: "AUTO", "bernoulli", "multinomial", "gaussian", "poisson", "gamma", "tweedie", "laplace", "quantile", "huber", "custom". Parameterized distributions can be passed as a dict, e.g., dict(type="tweedie", tweedie_power=1.5).
​
verbosity
string
default:"warn"
Verbosity of backend messages during training. One of None, "debug", "info", "warn", "error".

​
Methods

​
train

aml.train(
    x=None, y=None,
    training_frame=None,
    fold_column=None,
    weights_column=None,
    validation_frame=None,
    leaderboard_frame=None,
    blending_frame=None
)
Start the AutoML training process.
​
x
string[] | integer[]
Predictor column names or indices. When None, all columns except y are used.
​
y
string | integer
required
Response column name or index.
​
training_frame
H2OFrame
required
Training dataset.
​
fold_column
string
Column with pre-assigned cross-validation fold indices.
​
weights_column
string
Column containing per-row observation weights.
​
validation_frame
H2OFrame
Validation dataset. Only used when nfolds=0. When cross-validation is active, cross-validation metrics take precedence for early stopping.
​
leaderboard_frame
H2OFrame
Test dataset used to score the leaderboard. When omitted, cross-validation metrics are used.
​
blending_frame
H2OFrame
Frame used for training Stacked Ensemble metalearners (blending mode). Only used when nfolds=0.

​
predict

predictions = aml.predict(test_data)
Generate predictions using the current leader model.
​
test_data
H2OFrame
required
Data to score.
pred = aml.predict(test)

​
get_leaderboard

lb = aml.get_leaderboard(extra_columns="ALL")
Return the leaderboard with optional additional columns.
​
extra_columns
string | string[]
Extra metric columns to include. Pass "ALL" for all available metrics, or a list of specific metric names.

​
Properties

​
leaderboard

aml.leaderboard  # H2OFrame
The leaderboard H2OFrame ranking all models by the sort metric. Columns include model_id and the primary evaluation metric. 
aml.leaderboard.head(20)

​
leader

aml.leader  # H2OEstimator
The best-performing model from the leaderboard. This is a fully trained H2O model object on which you can call any model method (e.g., predict(), model_performance(), varimp()). 
best_model = aml.leader
perf = best_model.model_performance(test)

​
project_name

aml.project_name  # str
The project name associated with this AutoML run. Use this to retrieve the same run later with h2o.automl.get_automl(project_name).

​
Full example

import h2o
from h2o.automl import H2OAutoML

h2o.init()

# Load training and test data
train = h2o.import_file("https://s3.amazonaws.com/erin-data/higgs/higgs_train_10k.csv")
test = h2o.import_file("https://s3.amazonaws.com/erin-data/higgs/higgs_test_5k.csv")

y = "response"
x = [col for col in train.columns if col != y]

# Cast response to factor for classification
train[y] = train[y].asfactor()
test[y] = test[y].asfactor()

# Run AutoML for 30 seconds
aml = H2OAutoML(max_runtime_secs=30, seed=42)
aml.train(x=x, y=y, training_frame=train)

# Inspect leaderboard
print(aml.leaderboard.head(10))

# Evaluate best model on test set
perf = aml.leader.model_performance(test)
print(perf.auc())

# Predict with the leader model
pred = aml.predict(test)
pred.head()
To ensure reproducibility, set max_models instead of (or in addition to) max_runtime_secs. This guarantees the same models are trained regardless of hardware speed.

Build docs developers (and LLMs) love

Get started for freeTalk to us