Examples with NearestNeighbour org.knime.base.util.kdtree.NearestNeighbour used on opensource projects

Example 1 with NearestNeighbour

use of org.knime.base.util.kdtree.NearestNeighbour in project knime-core by knime.

the class KnnNodeModel method classify.

// returns a list where the first value if the winner class, and the
// following values are the class probabilities (if enabled)
private List<DataCell> classify(final DataRow row, final KDTree<DataCell> tree, final List<Integer> featureColumns, final Map<Integer, Integer> firstToSecond, final DataCell[] allClassValues) {
    double[] features = createQueryVector(row, featureColumns, firstToSecond);
    List<DataCell> output = new ArrayList<DataCell>();
    if (features == null) {
        for (int i = 0; i < 1 + allClassValues.length; i++) {
            output.add(DataType.getMissingCell());
        }
        return output;
    }
    HashMap<DataCell, MutableDouble> classWeights = new LinkedHashMap<DataCell, MutableDouble>();
    List<NearestNeighbour<DataCell>> nearestN = tree.getKNearestNeighbours(features, Math.min(m_settings.k(), tree.size()));
    for (NearestNeighbour<DataCell> n : nearestN) {
        MutableDouble count = classWeights.get(n.getData());
        if (count == null) {
            count = new MutableDouble(0);
            classWeights.put(n.getData(), count);
        }
        if (m_settings.weightByDistance()) {
            count.add(1 / n.getDistance());
        } else {
            count.inc();
        }
    }
    double winnerWeight = 0;
    double weightSum = 0;
    DataCell winnerCell = DataType.getMissingCell();
    for (Map.Entry<DataCell, MutableDouble> e : classWeights.entrySet()) {
        double weight = e.getValue().doubleValue();
        if (weight > winnerWeight) {
            winnerWeight = weight;
            winnerCell = e.getKey();
        }
        weightSum += weight;
    }
    // check if there are other classes with the same weight
    for (Map.Entry<DataCell, MutableDouble> e : classWeights.entrySet()) {
        double weight = e.getValue().doubleValue();
        if (weight == winnerWeight) {
            if (m_classDistribution.get(winnerCell).intValue() < m_classDistribution.get(e.getKey()).intValue()) {
                winnerCell = e.getKey();
            }
        }
    }
    output.add(winnerCell);
    if (m_settings.outputClassProbabilities()) {
        for (DataCell classVal : allClassValues) {
            MutableDouble v = classWeights.get(classVal);
            if (v == null) {
                output.add(new DoubleCell(0));
            // } else if (Double.isInfinite(v.doubleValue())) { // if distance to prototype is 0
            // output.add(new DoubleCell(1));
            } else {
                output.add(new DoubleCell(v.doubleValue() / weightSum));
            }
        }
    }
    return output;
}