Skip to main content
H2O-3’s explainability interface wraps a collection of explanation methods and visualizations into two functions: h2o.explain() for global explanations and h2o.explain_row() for row-level (local) explanations. Both work with individual models, lists of models, and AutoML objects. Visualizations use ggplot2 in R and matplotlib in Python.

​
Overview

h2o.explain() generates a suite of explanations for a model or group of models against a holdout dataset.For a single model:
  • Confusion Matrix (classification only)
  • Residual Analysis (regression only)
  • Variable Importance
  • Partial Dependence (PD) Plots
  • Individual Conditional Expectation (ICE) Plots
For a list of models or AutoML object: All of the above, plus:
  • Leaderboard
  • Variable Importance Heatmap
  • Model Correlation Heatmap
  • SHAP Summary of the top tree-based model
  • Partial Dependence Multi-model Plots

​
Quick start

import h2o
from h2o.automl import H2OAutoML
h2o.init()

# Import wine quality dataset
f = "https://h2o-public-test-data.s3.amazonaws.com/smalldata/wine/winequality-redwhite-no-BOM.csv"
df = h2o.import_file(f)
y = "quality"

splits = df.split_frame(ratios=[0.8], seed=1)
train = splits[0]
test = splits[1]

# Train AutoML for 60 seconds
aml = H2OAutoML(max_runtime_secs=60, seed=1)
aml.train(y=y, training_frame=train)

# Global explanations for all AutoML models
exa = aml.explain(test)

# Global explanations for the leader model only
exm = aml.leader.explain(test)

# Local explanation for the first row (0-indexed in Python)
aml.explain_row(test, row_index=0)
In R, the H2OExplanation object does not print automatically when saved to a variable. Call print(exa) or simply type exa to render it. In Python, the explanations print immediately even when saved to a variable.

​
Parameters

ParameterDescription
object (R)An H2O model, AutoML object, list of models, or H2OFrame with a model_id column (e.g. AutoML leaderboard).
newdata (R) / frame (Python)Holdout H2OFrame used for residual analysis, SHAP, and other explanations.
columnsColumn names to include in column-based explanations (PDP, ICE, SHAP). When specified, top_n_features is ignored.
top_n_featuresNumber of top features (ranked by variable importance) to include in column-based explanations. Defaults to 5.
include_explanationsGenerate only the listed explanation types. Mutually exclusive with exclude_explanations.
exclude_explanationsSkip the listed explanation types.
plot_overridesOverride arguments for individual plots (e.g., list(shap_summary_plot = list(top_n_features = 50))).
row_indexRow to explain (0-based in Python, 1-based in R). Used with explain_row().

​
Variable importance

The variable importance plot shows the relative contribution of each feature to the model’s predictions. 
# Variable importance for a single model
va_plot = model.varimp_plot()
Variable importance is not currently available for Stacked Ensemble models. When using explain() on an AutoML object with a Stacked Ensemble at the top, H2O-3 automatically shows variable importance for the top non-stacked base model instead.

​
Variable importance heatmap

The variable importance heatmap compares feature importance across multiple models. Categorical columns that were one-hot encoded (e.g., for Deep Learning or XGBoost) are summarized back to the original column level. Models and variables are ordered by similarity. 
# Variable importance heatmap across all AutoML models
va_heatmap = aml.varimp_heatmap()

# Use a subset of models sorted by MAE
va_heatmap = h2o.varimp_heatmap(aml.leaderboard.sort("mae").head(10))

​
SHAP summary plots

SHAP (SHapley Additive exPlanations) summary plots show each feature’s contribution for every row in the holdout dataset. The sum of feature contributions and the bias term equals the model’s raw prediction before applying the inverse link function. H2O-3 uses TreeSHAP, which is exact for tree-based models. Note that when features are correlated, TreeSHAP may assign higher importance to a feature that had no direct influence on a prediction. 
# SHAP summary plot for a single model
shap_plot = model.shap_summary_plot(test)

​
Row-level SHAP

Use shap_explain_row_plot() to see feature contributions for a single prediction. 
# SHAP for row index 0 (Python is 0-based)
shapr_plot = model.shap_explain_row_plot(test, row_index=0)

​
Partial dependence plots (PDP)

Partial dependence plots show the marginal effect of a single feature on the model’s predicted outcome, averaged over all other features. PDP assumes the target feature is independent of the remaining features. 
# PDP for a single model and feature
pd_plot = model.pd_plot(test, column="alcohol")

# PDP across multiple models (AutoML)
pd_multi = aml.pd_multi_plot(test, column="alcohol")

# PDP for a specific row becomes an ICE plot
pd_row = model.pd_plot(test, column="alcohol", row_index=0)

​
Individual Conditional Expectation (ICE) plots

ICE plots show the marginal effect of a feature for each individual instance, rather than the average across all instances. This reveals interaction effects that PDP can mask. 
# ICE plot for a feature
ice_plot = model.ice_plot(test, column="alcohol")

​
Residual analysis

Residual analysis plots fitted values vs. residuals on the test dataset. Randomly distributed residuals indicate a well-specified model. Patterns may indicate missing non-linear effects, heteroscedasticity, or autocorrelation.
If you see “striped” patterns in the residuals, this is typically caused by an integer-valued (rather than continuous) response variable, not a model issue.
# Residual analysis plot (regression models only)
ra_plot = model.residual_analysis_plot(test)

​
Model correlation heatmap

The model correlation heatmap shows how similar the predictions of different models are. For classification, it measures the frequency of identical predictions. Models are ordered by hierarchical clustering of their similarity. 
# Model correlation heatmap across all AutoML models
mc_plot = aml.model_correlation_heatmap(test)

# Use a subset of models
mc_plot = h2o.model_correlation_heatmap(
    aml.leaderboard.sort("mae").head(10), test
)

​
Selecting specific explanations

Use include_explanations or exclude_explanations to control which explanations are generated. Available explanation types:
NameScope
"leaderboard"AutoML and list of models only
"residual_analysis"Regression only
"confusion_matrix"Classification only
"varimp"Not available for Stacked Ensembles
"varimp_heatmap"Multi-model only
"model_correlation_heatmap"Multi-model only
"shap_summary"Single models only
"pdp"All
"ice"All
# Only generate variable importance and PDP
exm = model.explain(test, include_explanations=["varimp", "pdp"])

# Skip SHAP (can be slow for large datasets)
exm = model.explain(test, exclude_explanations=["shap_summary"])

Build docs developers (and LLMs) love

Get started for freeTalk to us