Class Stacking

JSAT/src/jsat/classifiers/boosting/Stacking.java:32–387

This provides an implementation of the Stacking ensemble method. Stacking learns several base classifiers and a top level classifier learns to predict the target based on the outputs of all the ensambled models. Historically a linear model (such as LinearBatch) is used, which translates to l

 * @author Edward Raff
 */
public class Stacking implements Classifier, Regressor
{

	private static final long serialVersionUID = -6173323872903232074L;
	private int folds;
    /**
     * The number of weights needed per model
     */
    private int weightsPerModel;
    private Classifier aggregatingClassifier;
    private List<Classifier> baseClassifiers;
    
    private Regressor aggregatingRegressor;
    private List<Regressor> baseRegressors;
    
    public static final int DEFAULT_FOLDS = 3;

    /**
     * Creates a new Stacking classifier
     * @param folds the number of cross validation folds for learning the base model
     * @param aggregatingClassifier the classifier used to merge the results of all the input classifiers
     * @param baseClassifiers the list of base classifiers to ensemble
     */
    public Stacking(int folds, Classifier aggregatingClassifier, List<Classifier> baseClassifiers)
    {
        if(baseClassifiers.size() < 2)
            throw new IllegalArgumentException("base classifiers must contain at least 2 elements, not " + baseClassifiers.size());
        setFolds(folds);
        this.aggregatingClassifier = aggregatingClassifier;
        this.baseClassifiers = baseClassifiers;
        
        boolean allRegressors = aggregatingClassifier instanceof Regressor;
        for(Classifier cl : baseClassifiers)
            if(!(cl instanceof Regressor))
                allRegressors = false;
        
        if(allRegressors)
        {
            aggregatingRegressor = (Regressor) aggregatingClassifier;
            baseRegressors = (List) baseClassifiers;//ugly type easure exploitation... 
        }
    }
    
    /**
     * Creates a new Stacking classifier
     * @param folds the number of cross validation folds for learning the base model
     * @param aggregatingClassifier the classifier used to merge the results of all the input classifiers
     * @param baseClassifiers the array of base classifiers to ensemble
     */
    public Stacking(int folds, Classifier aggregatingClassifier, Classifier... baseClassifiers)
    {
        this(folds, aggregatingClassifier, Arrays.asList(baseClassifiers));
    }
    
    /**
     * Creates a new Stacking classifier that uses {@value #DEFAULT_FOLDS} folds of cross validation
     * @param aggregatingClassifier the classifier used to merge the results of all the input classifiers
     * @param baseClassifiers the list of base classifiers to ensemble