What are the most critical parameters to tweak for each specific H2O model type?

Machine Learning
,
1

To get strong performance from H2O models, it’s important to tune the parameters that matter most for each algorithm. Below is a practical guide to the key “levers” for each major H2O model type, based on common Sparkflows usage patterns.

1. H2O Gradient Boosting Models (GBM & XGBoost)

These models build trees sequentially, with each new tree correcting the mistakes of the previous ones. They are often the strongest performers.

Critical parameters:

  • learnRate: Controls how aggressively the model learns. Lower values (e.g., 0.01) are more stable and usually require higher ntrees.

  • ntrees: Number of trees. Works hand-in-hand with learnRate.

  • maxDepth: Limits tree complexity. Deeper trees capture complex interactions but increase overfitting risk.

  • scalePosWeight (XGBoost): Essential for imbalanced classification. Set this to the ratio of negative to positive samples.

  • regAlpha & regLambda (XGBoost): L1 and L2 regularization to control model complexity.

  • treeMethod (XGBoost): Use hist or approx for large datasets to improve training speed.

  • withContributions: Enables SHAP values for per-prediction feature explanations.

2. H2O Distributed Random Forest (DRF)

DRF builds many independent trees and averages their predictions, making it robust and easy to tune.

Critical parameters:

  • ntrees: More trees generally improve performance until diminishing returns.

  • mtries: Number of features considered at each split. Controls diversity among trees.

  • sampleRate: Percentage of rows each tree sees. Helps balance bias vs variance.

  • maxDepth: Prevents trees from growing too complex.

  • balanceClasses: Important for imbalanced classification problems.

  • withContributions: Enables SHAP-based interpretability.

3. H2O Deep Learning (Neural Networks)

Best suited for capturing complex, non-linear patterns but requires careful regularization.

Critical parameters:

  • hidden: Defines the network architecture (e.g., 200,200 for two hidden layers).

  • activation: Rectifier (ReLU), Tanh, or Dropout variants for better generalization.

  • l1 & l2: Regularization terms to prevent overfitting.

  • epochs: Number of full passes over the data.

  • rate / adaptiveRate: Controls learning dynamics during training.

4. H2O Generalized Linear Model (GLM)

A fast, interpretable baseline model widely used in production.

Critical parameters:

  • family: Must match the prediction task (bernoulli, multinomial, gaussian, etc.).

  • lambdaSearch: Automatically finds the optimal regularization strength.

  • solver: IRLSM for smaller datasets, L_BFGS for large or high-dimensional data.

  • computePValues: Enables statistical significance testing for model coefficients.

5. H2O Isolation Forest (Anomaly Detection)

Designed specifically for identifying rare and unusual observations.

Critical parameters:

  • contamination: Estimated proportion of anomalies in the dataset; defines the anomaly threshold.

  • ntrees: More trees increase stability of anomaly scores.

  • sampleSize: Controls how much data each tree sees.

6. Unsupervised Models

KMeans

  • k: Number of clusters (primary tuning lever).

  • estimateK: Lets H2O automatically search for an optimal number of clusters.

PCA & GLRM

  • k: Number of latent components.

  • transform (PCA): Use STANDARDIZE to ensure features are comparable.

  • regularizationX / regularizationY (GLRM): Controls complexity of latent factors.

Automating Tuning with Grid Search

Instead of manually guessing parameter values, H2O Grid Search allows systematic tuning:

  • paramKeys: Parameters to tune (e.g., ntrees, maxDepth).

  • paramValues: Candidate values for each parameter.

  • gridStrategy: Cartesian tests all combinations to find the best-performing model.

Bottom line:
Each H2O algorithm has a small set of high-impact parameters. Focusing on those—and combining them with Grid Search and SHAP explainability—delivers the biggest gains in both model performance and trust.