Examples with ImpurityCalculator org.apache.spark.mllib.tree.impurity.ImpurityCalculator used on opensource projects

Example 1 with ImpurityCalculator

use of org.apache.spark.mllib.tree.impurity.ImpurityCalculator in project jpmml-sparkml by jpmml.

the class TreeModelUtil method encodeNode.

public static Node encodeNode(org.apache.spark.ml.tree.Node node, PredicateManager predicateManager, Map<FieldName, Set<String>> parentFieldValues, MiningFunction miningFunction, Schema schema) {
    if (node instanceof InternalNode) {
        InternalNode internalNode = (InternalNode) node;
        Map<FieldName, Set<String>> leftFieldValues = parentFieldValues;
        Map<FieldName, Set<String>> rightFieldValues = parentFieldValues;
        Predicate leftPredicate;
        Predicate rightPredicate;
        Split split = internalNode.split();
        Feature feature = schema.getFeature(split.featureIndex());
        if (split instanceof ContinuousSplit) {
            ContinuousSplit continuousSplit = (ContinuousSplit) split;
            double threshold = continuousSplit.threshold();
            if (feature instanceof BooleanFeature) {
                BooleanFeature booleanFeature = (BooleanFeature) feature;
                if (threshold != 0.5d) {
                    throw new IllegalArgumentException();
                }
                leftPredicate = predicateManager.createSimplePredicate(booleanFeature, SimplePredicate.Operator.EQUAL, booleanFeature.getValue(0));
                rightPredicate = predicateManager.createSimplePredicate(booleanFeature, SimplePredicate.Operator.EQUAL, booleanFeature.getValue(1));
            } else {
                ContinuousFeature continuousFeature = feature.toContinuousFeature();
                String value = ValueUtil.formatValue(threshold);
                leftPredicate = predicateManager.createSimplePredicate(continuousFeature, SimplePredicate.Operator.LESS_OR_EQUAL, value);
                rightPredicate = predicateManager.createSimplePredicate(continuousFeature, SimplePredicate.Operator.GREATER_THAN, value);
            }
        } else if (split instanceof CategoricalSplit) {
            CategoricalSplit categoricalSplit = (CategoricalSplit) split;
            double[] leftCategories = categoricalSplit.leftCategories();
            double[] rightCategories = categoricalSplit.rightCategories();
            if (feature instanceof BinaryFeature) {
                BinaryFeature binaryFeature = (BinaryFeature) feature;
                SimplePredicate.Operator leftOperator;
                SimplePredicate.Operator rightOperator;
                if (Arrays.equals(TRUE, leftCategories) && Arrays.equals(FALSE, rightCategories)) {
                    leftOperator = SimplePredicate.Operator.EQUAL;
                    rightOperator = SimplePredicate.Operator.NOT_EQUAL;
                } else if (Arrays.equals(FALSE, leftCategories) && Arrays.equals(TRUE, rightCategories)) {
                    leftOperator = SimplePredicate.Operator.NOT_EQUAL;
                    rightOperator = SimplePredicate.Operator.EQUAL;
                } else {
                    throw new IllegalArgumentException();
                }
                String value = ValueUtil.formatValue(binaryFeature.getValue());
                leftPredicate = predicateManager.createSimplePredicate(binaryFeature, leftOperator, value);
                rightPredicate = predicateManager.createSimplePredicate(binaryFeature, rightOperator, value);
            } else if (feature instanceof CategoricalFeature) {
                CategoricalFeature categoricalFeature = (CategoricalFeature) feature;
                FieldName name = categoricalFeature.getName();
                List<String> values = categoricalFeature.getValues();
                if (values.size() != (leftCategories.length + rightCategories.length)) {
                    throw new IllegalArgumentException();
                }
                final Set<String> parentValues = parentFieldValues.get(name);
                com.google.common.base.Predicate<String> valueFilter = new com.google.common.base.Predicate<String>() {

                    @Override
                    public boolean apply(String value) {
                        if (parentValues != null) {
                            return parentValues.contains(value);
                        }
                        return true;
                    }
                };
                List<String> leftValues = selectValues(values, leftCategories, valueFilter);
                List<String> rightValues = selectValues(values, rightCategories, valueFilter);
                leftFieldValues = new HashMap<>(parentFieldValues);
                leftFieldValues.put(name, new HashSet<>(leftValues));
                rightFieldValues = new HashMap<>(parentFieldValues);
                rightFieldValues.put(name, new HashSet<>(rightValues));
                leftPredicate = predicateManager.createSimpleSetPredicate(categoricalFeature, leftValues);
                rightPredicate = predicateManager.createSimpleSetPredicate(categoricalFeature, rightValues);
            } else {
                throw new IllegalArgumentException();
            }
        } else {
            throw new IllegalArgumentException();
        }
        Node result = new Node();
        Node leftChild = encodeNode(internalNode.leftChild(), predicateManager, leftFieldValues, miningFunction, schema).setPredicate(leftPredicate);
        Node rightChild = encodeNode(internalNode.rightChild(), predicateManager, rightFieldValues, miningFunction, schema).setPredicate(rightPredicate);
        result.addNodes(leftChild, rightChild);
        return result;
    } else if (node instanceof LeafNode) {
        LeafNode leafNode = (LeafNode) node;
        Node result = new Node();
        switch(miningFunction) {
            case REGRESSION:
                {
                    String score = ValueUtil.formatValue(node.prediction());
                    result.setScore(score);
                }
                break;
            case CLASSIFICATION:
                {
                    CategoricalLabel categoricalLabel = (CategoricalLabel) schema.getLabel();
                    int index = ValueUtil.asInt(node.prediction());
                    result.setScore(categoricalLabel.getValue(index));
                    ImpurityCalculator impurityCalculator = node.impurityStats();
                    result.setRecordCount((double) impurityCalculator.count());
                    double[] stats = impurityCalculator.stats();
                    for (int i = 0; i < stats.length; i++) {
                        ScoreDistribution scoreDistribution = new ScoreDistribution(categoricalLabel.getValue(i), stats[i]);
                        result.addScoreDistributions(scoreDistribution);
                    }
                }
                break;
            default:
                throw new UnsupportedOperationException();
        }
        return result;
    } else {
        throw new IllegalArgumentException();
    }
}