Skip to main content
Distributed Random Forest (DRF) is a powerful classification and regression algorithm. Given a dataset, DRF generates a forest of decision trees rather than a single tree. Each tree is a weak learner built on a random subset of rows and columns (controlled by sample_rate and mtries). The final prediction averages all trees, reducing variance without increasing bias. DRF supports both classification (majority vote) and regression (average prediction), per-row observation weights, N-fold cross-validation, and produces calibrated probabilities. MOJO Support: DRF fully supports importing and exporting MOJOs.

​
Extremely Randomized Trees (XRT)

H2O-3 supports Extremely Randomized Trees via histogram_type="Random". Unlike standard DRF where the best threshold is found for each candidate feature, XRT draws split thresholds at random, further reducing variance at the cost of a small bias increase. Enable it by setting: 
drf = H2ORandomForestEstimator(histogram_type="Random")

​
Key Parameters

​
ntrees
int
default:"50"
Number of trees in the forest. More trees reduce variance but increase training time. Unlike GBM, DRF does not overfit badly with more trees — test accuracy plateaus rather than degrades.
​
max_depth
int
default:"20"
Maximum tree depth. DRF uses deeper trees than GBM by default (20 vs 5) because averaging reduces the overfitting risk of individual deep trees.
​
mtries
int
default:"-1"
Number of columns randomly sampled at each split. -1 uses the square root of the number of columns for classification and p/3 for regression (standard random forest defaults). -2 uses all features (equivalent to bagging). Any value >= 1 sets the exact count.
​
sample_rate
float
default:"0.632"
Row subsampling rate per tree (without replacement). The default 0.632 is the classic bootstrap approximation. Reducing this value adds more randomness and can improve generalization on noisy data.
​
min_rows
float
default:"1.0"
Minimum observations required in a leaf node. Increase to smooth predictions on small datasets. R parameter name: node_size.
​
nbins
int
default:"20"
Number of histogram bins for numerical columns. More bins capture finer split points but increase memory and training time.
​
nbins_cats
int
default:"1024"
Maximum bins for categorical columns. Higher values allow finer splits on high-cardinality categoricals.
​
histogram_type
str
default:"AUTO"
Split-finding strategy: "AUTO", "UniformAdaptive", "UniformRobust", "Random" (XRT), "QuantilesGlobal", "RoundRobin".
​
binomial_double_trees
bool
default:"False"
Binary classification only. Build twice as many trees (one per class). Can improve accuracy at the cost of doubled training time.

​
Code Examples

import h2o
from h2o.estimators.random_forest import H2ORandomForestEstimator

h2o.init()

train = h2o.import_file("https://s3.amazonaws.com/h2o-public-test-data/smalldata/higgs/higgs_train_5k.csv")
test  = h2o.import_file("https://s3.amazonaws.com/h2o-public-test-data/smalldata/higgs/higgs_test_5k.csv")

y = "response"
x = [c for c in train.columns if c != y]
train[y] = train[y].asfactor()
test[y]  = test[y].asfactor()

drf = H2ORandomForestEstimator(
    ntrees=100,
    max_depth=20,
    mtries=-1,          # sqrt(p) for classification
    sample_rate=0.632,
    stopping_rounds=5,
    stopping_metric="AUC",
    seed=42
)
drf.train(x=x, y=y, training_frame=train, validation_frame=test)

# Variable importance
vi = drf.varimp(use_pandas=True)
print(vi.head(10))

print(drf.auc(valid=True))

​
Variable Importance

DRF computes variable importance based on mean decrease in squared error across all splits that use each feature. Access it as follows:
Python
# As a data frame (pandas)
vi = drf.varimp(use_pandas=True)
print(vi[["variable", "relative_importance", "scaled_importance", "percentage"]].head(20))

# Plot
drf.varimp_plot()
R
# Print
h2o.varimp(drf)

# Plot
h2o.varimp_plot(drf)
Variable importance in DRF is computed from the training data. For a more reliable estimate of feature relevance, use SHAP (Shapley) contributions via drf.predict_contributions(test) in Python or h2o.predict_contributions(drf, test) in R. SHAP provides per-row, per-feature attribution rather than a global average.

​
DRF vs GBM

PropertyDRFGBM
Training strategyParallel (bagging)Sequential (boosting)
Default max_depth205
Default sample_rate0.6321.0
Overfitting risk with many treesLow (variance reduces)Moderate (use early stopping)
Tuning effortLowHigher
Typical AUCGood baselineSlightly higher ceiling
Training timeFaster (fully parallel)Slower (sequential)
DRF is an excellent first-pass model and a strong baseline with minimal tuning. GBM and XGBoost typically achieve slightly higher accuracy with proper hyperparameter search.

Build docs developers (and LLMs) love

Get started for freeTalk to us