06_​H2O_​GBM_​parameter_​optimization

0 ×

This tutorial shows how to train multiple H2O Models in KNIME using parameter optimization (grid search) and extract the optimal algorithm settings for the training of the final model. We will train Gradient Boosting Machines for binominal classification using a grid of two different GBM parameters.

1. Prepare:
Load and Import data to H2O.

2. Optimization:
To train models with parameter optimization, we create a Loop using the KNIME Node "Parameter Optimization Loop Start" (Analytics - Mining). In this Nodes' settings we can define the optimization grid: For this example we will optimize the GBM algorithm parameters "Number of trees" and "Max tree depth". We use brute force optimization, meaning that there will be as many iteration as there are parameter combinations defined in the Parameter Optimization Loop Start Node. The "Loop End" Node collects the scored metrics of all optimization loop iterations. In order to extract the optimal algorithm parameters, we sort the collected rows by several metrics and filter the top row.

3. Learn Models, do prediction and scoring in Parameter Optimization Loop:
For each combination of parameters, a GBM Model is build by H2O using the "Number of Trees" and "Max tree depth" parameters of the corresponding loop iteration and the model accuracy metrics are scored.

4. Train final model
Finally, we use the optimal parameters to predict new data.

Nodes

• File Reader2 ×
• H2O Gradient Boosting Machine Learner2 ×
• H2O Predictor (Classification)2 ×
• Table to H2O2 ×
• Cross Joiner1 ×
• H2O Binomial Scorer1 ×
• H2O Local Context1 ×
• H2O Partitioning1 ×
• Loop End1 ×
• Parameter Optimization Loop Start1 ×
• Row Filter1 ×
• Sorter1 ×
• Table Row to Variable1 ×
• Variable to Table Row1 ×

Extensions

• FeatureKNIME Base nodes
• FeatureKNIME H2O Machine Learning Integration
• FeatureKNIME Optimization extension

Links

• H2O documentation

  http://docs.h2o.ai/h2o/latest-stable/h2o-docs/index.html