Skip to main content
H2O’s AutoML automates the supervised machine learning workflow. It trains several models — cross-validated by default — across the following algorithms: XGBoost, GBM, GLM, DRF, XRT (Extremely Randomized Trees), and Deep Learning. It also applies hyperparameter optimization on XGBoost, GBM, and Deep Learning, and finally trains multiple Stacked Ensemble models on top of the best base learners.
AutoML is available in Python (H2OAutoML), R (h2o.automl()), and the H2O Flow web UI.

​
How AutoML Works

1

Data ingestion

Point AutoML at a training frame and identify the response column. Everything else is optional.
2

Model training

AutoML sequentially trains individual models (GLM, GBM, XGBoost, DRF, XRT, Deep Learning) and performs random grid searches with hyperparameter optimization for the strongest algorithms.
3

Stacked Ensembles

After base model training, AutoML builds two Stacked Ensemble models: one using all base models and one using only the best model from each algorithm family.
4

Leaderboard ranking

All models are ranked on the leaderboard by cross-validation AUC (binary), mean per-class error (multinomial), or deviance (regression) — or by a custom sort_metric you specify.

​
Parameters

​
Required Parameters

​
y
str
required
The name (or index) of the response column. For binary classification, this column must be a factor.
​
training_frame
H2OFrame
required
The dataset used to build the models. At least one of max_runtime_secs or max_models must also be set (or AutoML defaults to 1 hour).
​
max_runtime_secs
int
default:"3600"
Maximum wall-clock time (seconds) for the entire AutoML run. Dynamically defaults to 3600 if neither max_runtime_secs nor max_models is specified.
​
max_models
int
default:"None"
Maximum number of models to train (excluding Stacked Ensembles). Set this for reproducible runs — all models are then trained to convergence rather than being cut off by a time budget.

​
Key Optional Parameters

​
nfolds
int
default:"-1"
Number of cross-validation folds. -1 lets AutoML decide (typically 5-fold CV or blending mode depending on data size). Set to 0 to disable CV (also disables Stacked Ensembles).
​
seed
int
default:"None"
Random seed for reproducibility. Reproducibility requires using max_models (not max_runtime_secs) and excluding DeepLearning (which is non-deterministic by default).
​
exclude_algos
List[str]
default:"None"
List of algorithms to skip. Mutually exclusive with include_algos. Example: ["GLM", "DeepLearning"]. Available values: "DRF", "GLM", "XGBoost", "GBM", "DeepLearning", "StackedEnsemble".
​
include_algos
List[str]
default:"None"
Allowlist of algorithms. Mutually exclusive with exclude_algos. Same values as above.
​
sort_metric
str
default:"AUTO"
Metric used to rank the leaderboard. AUTO defaults to AUC for binary, mean_per_class_error for multinomial, and deviance for regression. Other options: "logloss", "MSE", "RMSE", "MAE".
​
verbosity
str
default:"warn"
Backend logging verbosity. One of "debug", "info", "warn". Set to "info" to see per-model progress during training.

​
Code Examples

import h2o
from h2o.automl import H2OAutoML

h2o.init()

# Load 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")

# Response column must be a factor for classification
y = "response"
train[y] = train[y].asfactor()
test[y]  = test[y].asfactor()

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

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

# View the leaderboard
lb = aml.leaderboard
lb.head(rows=lb.nrows)  # print full leaderboard

# Evaluate the leader model
perf = aml.leader.model_performance(test)
print(perf.auc())

​
Accessing and Interpreting the Leaderboard

The leaderboard is an H2OFrame sorted by the sort_metric. Each row represents one trained model.
Python
# Full leaderboard
lb = aml.leaderboard
lb.head(rows=lb.nrows)

# Get the best model
best_model = aml.leader

# Get a specific model by ID
m = h2o.get_model(lb["model_id"][2, 0])
R
# Full leaderboard
lb <- aml@leaderboard
print(lb, n = nrow(lb))

# Leader
leader <- aml@leader

# Get model by ID
m <- h2o.getModel(as.character(lb[3, "model_id"]))
Stacked Ensemble models typically appear at the top of the leaderboard. If you need a single interpretable model, look for the best GLM entry or the best GBM/XGBoost entry — accessible via aml.get_best_model(algorithm="glm").

​
Getting the Best Model by Algorithm

Python
# Best model overall
best = aml.leader

# Best model of a specific algorithm
best_gbm = aml.get_best_model(algorithm="gbm")
best_xgb = aml.get_best_model(algorithm="xgboost")
best_glm = aml.get_best_model(algorithm="glm")

​
Explainability

H2O AutoML integrates with H2O’s explainability framework. Call explain() on the AutoML object to generate an automated report covering variable importance, SHAP values, partial dependence plots, and model correlation.
Python
# Explain all AutoML models on the test set
exa = aml.explain(test)

# Explain only the leader model
exm = aml.leader.explain(test)
R
# Explain AutoML results
exa <- h2o.explain(aml, test)

# Explain leader model
exm <- h2o.explain(aml@leader, test)
explain() returns an object with individual plots (variable importance heatmap, model correlation heatmap, SHAP summary, PDP/ICE plots). In a Jupyter notebook, plots render inline automatically.

Build docs developers (and LLMs) love

Get started for freeTalk to us