# Overview

In this tutorial, we will discuss ensemble learning with a focus on a type of ensemble learning called stacking or Super Learning. In this tutorial, we present an H2O implementation of the Super Learner algorithm (aka. Stacking, Stacked Ensembles).

Following the introduction to ensemble learning, we will dive into a hands-on code demo of the h2oEnsemble R package. Note that as of H2O 3.10.3.1, Stacked Ensembles are now available as part of base H2O. The documentation for H2O Stacked Ensembles, including R and Python code examples, can be found here. The h2oEnsemble R package is the predecessor to the base H2O implementation and although this package will contibue to be supported, new development efforts will be focused on the native H2O version of stacked ensembles, and for new projects we’d recommend using native H2O.

H2O World slides accompanying this tutorial are here.

The GitHub page for the ensembles is here.

# What is Ensemble Learning?

Ensemble machine learning methods use multiple learning algorithms to obtain better predictive performance than could be obtained from any of the constituent learning algorithms.

Many of the popular modern machine learning algorithms are actually ensembles. For example, Random Forest and Gradient Boosting Machine are both ensemble learners.

Common types of ensembles:
Bagging
Boosting
Stacking

## Bagging

Bootstrap aggregating, or bagging, is an ensemble method designed to improve the stability and accuracy of machine learning algorithms. It reduces variance and helps to avoid overfitting. Bagging is a special case of the model averaging approach and is relatively robust against noisy data and outliers.

One of the most well known bagging ensembles is the Random Forest algorithm, which applies bagging to decision trees.

## Boosting

Boosting is an ensemble method designed to reduce bias and variance. A boosting algorithm iteratively learns weak classifiers and adds them to a final strong classifier.

After a weak learner is added, the data is reweighted: examples that are misclassified gain weight and examples that are classified correctly lose weight. Thus, future weak learners focus more on the examples that previous weak learners misclassified. This causes boosting methods to be not very robust to noisy data and outliers.

Both bagging and boosting are ensembles that take a collection of weak learners and forms a single, strong learner.

## Stacking / Super Learning

Stacking is a broad class of algorithms that involves training a second-level “metalearner” to ensemble a group of base learners. The type of ensemble learning implemented in H2O is called “super learning”, “stacked regression” or “stacking.” Unlike bagging and boosting, the goal in stacking is to ensemble strong, diverse sets of learners together.

### Some Background

Leo Breiman, known for his work on classification and regression trees and the creator of the Random Forest algorithm, formalized stacking in his 1996 paper, “Stacked Regressions”. Although the idea originated with David Wolpert in 1992 under the name “Stacked Generalization”, the modern form of stacking that uses internal k-fold cross-validation was Dr. Breiman’s contribution.

However, it wasn’t until 2007 that the theoretical background for stacking was developed, which is when the algorithm took on the name, “Super Learner”. Until this time, the mathematical reasons for why stacking worked were unknown and stacking was considered a “black art.” The Super Learner algorithm learns the optimal combination of the base learner fits. In an article titled, “Super Learner”, by Mark van der Laan et al., proved that the Super Learner ensemble represents an asymptotically optimal system for learning.

### Super Learner Algorithm

Here is an outline of the tasks involved in training and testing a Super Learner ensemble.

#### Set up the ensemble

• Specify a list of L base algorithms (with a specific set of model parameters).
• Specify a metalearning algorithm.

#### Train the ensemble

• Train each of the L base algorithms on the training set.
• Perform k-fold cross-validation on each of these learners and collect the cross-validated predicted values from each of the L algorithms.
• The N cross-validated predicted values from each of the L algorithms can be combined to form a new N x L matrix. This matrix, along wtih the original response vector, is called the “level-one” data. (N = number of rows in the training set)
• Train the metalearning algorithm on the level-one data.
• The “ensemble model” consists of the L base learning models and the metalearning model, which can then be used to generate predictions on a test set.

#### Predict on new data

• To generate ensemble predictions, first generate predictions from the base learners.
• Feed those predictions into the metalearner to generate the ensemble prediction.

# H2O Ensemble: Super Learning in H2O

H2O Ensemble has been implemented as a stand-alone R package called h2oEnsemble. The package is an extension to the h2o R package that allows the user to train an ensemble in the H2O cluster using any of the supervised machine learning algorithms H2O. As in the h2o R package, all of the actual computation in h2oEnsemble is performed inside the H2O cluster, rather than in R memory.

The main computational tasks in the Super Learner ensemble algorithm are the training and cross-validation of the base learners and metalearner. Therefore, implementing the “plumbing” of the ensemble in R (rather than in Java) does not incur a loss of performance. All training and data processing are performed in the high-performance H2O cluster.

H2O Ensemble currently supports regression and binary classification. Multi-class support will be added in a future release.

## Install H2O Ensemble

To install the h2oEnsemble package, you just need to follow the installation instructions on the README file, also documented here for convenience.

### H2O R Package

First you need to install the H2O R package if you don’t already have it installed. The R installation instructions are at: http://h2o.ai/download

### H2O Ensemble R Package

The recommended way of installing the h2oEnsemble R package is directly from GitHub using the devtools package (however, H2O World tutorial attendees should install the package from the provided USB stick).

## Higgs Demo

This is an example of binary classification using the h2o.ensemble function, which is available in h2oEnsemble. This demo uses a subset of the HIGGS dataset, which has 28 numeric features and a binary response. The machine learning task in this example is to distinguish between a signal process which produces Higgs bosons (Y = 1) and a background process which does not (Y = 0). The dataset contains approximately the same number of positive vs negative examples. In other words, this is a balanced, rather than imbalanced, dataset.

If run from plain R, execute R in the directory of this script. If run from RStudio, be sure to setwd() to the location of this script. h2o.init() starts H2O in R’s current working directory. h2o.importFile() looks for files from the perspective of where H2O was started.

### Load Data into H2O Cluster

First, import a sample binary outcome train and test set into the H2O cluster.

For binary classification, the response should be encoded as a factor type (also known as the enum type in Java or categorial in Python Pandas). The user can specify column types in the h2o.importFile command, or you can convert the response column as follows:

### Specify Base Learners & Metalearner

For this example, we will use the default base learner library for h2o.ensemble, which includes the default H2O GLM, Random Forest, GBM and Deep Neural Net (all using default model parameter values). We will also use the default metalearner, the H2O GLM.

### Train an Ensemble

Train the ensemble (using 5-fold internal CV) to generate the level-one data. Note that more CV folds will take longer to train, but should increase performance.

### Evaluate Model Performance

Since the response is binomial, we can use Area Under the ROC Curve (AUC) to evaluate the model performance. Compute test set performance, and sort by AUC (the default metric that is printed for a binomial classification):

We can compare the performance of the ensemble to the performance of the individual learners in the ensemble.

So we see the best individual algorithm in this group is the GBM with a test set AUC of 0.778, as compared to 0.782 for the ensemble. At first thought, this might not seem like much, but in many industries like medicine or finance, this small advantage can be highly valuable.

To increase the performance of the ensemble, we have several options. One of them is to increase the number of internal cross-validation folds using the cvControl argument. The other options are to change the base learner library or the metalearning algorithm.

Note that the ensemble results above are not reproducible since h2o.deeplearning is not reproducible when using multiple cores, and we did not set a seed for h2o.randomForest.wrapper.

If we want to evaluate the model by a different metric, say “MSE”, then we can pass that metric to the print method for and ensemble performance object as follows:

### Predict

If you actually need to generate the predictions (instead of looking only at model performance), you can use the predict() function with a test set. Generate predictions on the test set and store as an H2O Frame:

If you need to bring the predictions back into R memory for futher processing, you can convert pred to a local R data.frame as follows:

The predict method for an h2o.ensemble fit will return a list of two objects. The pred$pred object contains the ensemble predictions, and pred$basepred is a matrix of predictions from each of the base learners. In this particular example where we used four base learners, the pred\$basepred matrix has four columns. Keeping the base learner predictions around is useful for model inspection and will allow us to calculate performance of each of the base learners on the test set (for comparison to the ensemble).

### Specifying new learners

Now let’s try again with a more extensive set of base learners. The h2oEnsemble packages comes with four functions by default that can be customized to use non-default parameters.

Here is an example of how to generate a custom learner wrappers:

Let’s grab a subset of these learners for our base learner library and re-train the ensemble.

### Customized base learner library

Train with new library:

Evaluate the test set performance:

So what happens to the ensemble if we remove some of the weaker learners? Let’s remove the GLM and DL from the learner library and see what happens…

Here is a more stripped down version of the base learner library used above:

Again re-train the ensemble and evaluate the performance:

At first thought, you may assume that removing less performant models would increase the perforamnce of the ensemble. However, each learner has it’s own unique contribution to the ensemble and the added diversity among learners usually improves performance. The Super Learner algorithm learns the optimal way of combining all these learners together in a way that is superior to other combination/blending methods.

### Stacking Existing Model Sets

You can also use an existing (cross-validated) list of H2O models as the starting point and use the h2o.stack() function to ensemble them together via a specified metalearner. The base models must have been trained on the same dataset with same response and for cross-validation, must have all used the same folds.

An example follows. As above, start up the H2O cluster and load the training and test data.

Cross-validate and train a handful of base learners and then use the h2o.stack() function to create the ensemble: