Examples with IBoundingBoxView com.amazon.randomcutforest.tree.IBoundingBoxView used on opensource projects

Example 6 with IBoundingBoxView

use of com.amazon.randomcutforest.tree.IBoundingBoxView in project random-cut-forest-by-aws by aws.

the class AnomalyScoreVisitorTest method testGetProbabilityOfSeparation.

@Test
public void testGetProbabilityOfSeparation() {
    float[] minPoint = { 0.0f, 0.0f, 0.0f };
    float[] maxPoint = { 1.0f, 2.0f, 3.0f };
    IBoundingBoxView boundingBox = new BoundingBox(minPoint);
    boundingBox = boundingBox.getMergedBox(maxPoint);
    float[] point = { 0.5f, 0.5f, 0.5f };
    int sampleSize = 2;
    AnomalyScoreVisitor visitor = new AnomalyScoreVisitor(point, sampleSize);
    double p = visitor.getProbabilityOfSeparation(boundingBox);
    assertThat(p, closeTo(0.0, EPSILON));
    assertTrue(visitor.coordInsideBox[0]);
    assertTrue(visitor.coordInsideBox[1]);
    assertTrue(visitor.coordInsideBox[2]);
    visitor = new AnomalyScoreVisitor(point, sampleSize);
    visitor.coordInsideBox[1] = visitor.coordInsideBox[2] = true;
    p = visitor.getProbabilityOfSeparation(boundingBox);
    assertThat(p, closeTo(0.0, EPSILON));
    assertTrue(visitor.coordInsideBox[0]);
    assertTrue(visitor.coordInsideBox[1]);
    assertTrue(visitor.coordInsideBox[2]);
    point = new float[] { 2.0f, 0.5f, 0.5f };
    visitor = new AnomalyScoreVisitor(point, sampleSize);
    p = visitor.getProbabilityOfSeparation(boundingBox);
    assertThat(p, closeTo(1.0 / (2.0 + 2.0 + 3.0), EPSILON));
    assertFalse(visitor.coordInsideBox[0]);
    assertTrue(visitor.coordInsideBox[1]);
    assertTrue(visitor.coordInsideBox[2]);
    visitor = new AnomalyScoreVisitor(point, sampleSize);
    visitor.coordInsideBox[1] = visitor.coordInsideBox[2] = true;
    p = visitor.getProbabilityOfSeparation(boundingBox);
    assertThat(p, closeTo(1.0 / (2.0 + 2.0 + 3.0), EPSILON));
    assertFalse(visitor.coordInsideBox[0]);
    assertTrue(visitor.coordInsideBox[1]);
    assertTrue(visitor.coordInsideBox[2]);
    point = new float[] { 0.5f, -3.0f, 4.0f };
    visitor = new AnomalyScoreVisitor(point, sampleSize);
    p = visitor.getProbabilityOfSeparation(boundingBox);
    assertThat(p, closeTo((3.0 + 1.0) / (1.0 + 5.0 + 4.0), EPSILON));
    assertTrue(visitor.coordInsideBox[0]);
    assertFalse(visitor.coordInsideBox[1]);
    assertFalse(visitor.coordInsideBox[2]);
    visitor = new AnomalyScoreVisitor(point, sampleSize);
    visitor.coordInsideBox[0] = true;
    p = visitor.getProbabilityOfSeparation(boundingBox);
    assertThat(p, closeTo((3.0 + 1.0) / (1.0 + 5.0 + 4.0), EPSILON));
    assertTrue(visitor.coordInsideBox[0]);
    assertFalse(visitor.coordInsideBox[1]);
    assertFalse(visitor.coordInsideBox[2]);
}

Example 7 with IBoundingBoxView

use of com.amazon.randomcutforest.tree.IBoundingBoxView in project random-cut-forest-by-aws by aws.

the class AbstractAttributionVisitor method accept.

/**
 * Update the anomaly score based on the next step of the tree traversal.
 *
 * @param node        The current node in the tree traversal
 * @param depthOfNode The depth of the current node in the tree
 */
@Override
public void accept(INodeView node, int depthOfNode) {
    if (pointInsideBox) {
        return;
    }
    IBoundingBoxView smallBox;
    if (hitDuplicates || ignoreLeaf) {
        // use the sibling bounding box to represent counterfactual "what if point & the
        // candidate near neighbor
        // had not been inserted in the tree"
        shadowBox = shadowBox == null ? node.getSiblingBoundingBox(pointToScore) : shadowBox.getMergedBox(node.getSiblingBoundingBox(pointToScore));
        smallBox = shadowBox;
    } else {
        smallBox = node.getBoundingBox();
    }
    IBoundingBoxView largeBox = smallBox.getMergedBox(pointToScore);
    updateRangesForScoring(smallBox, largeBox);
    double probOfCut = sumOfDifferenceInRange / sumOfNewRange;
    // if leaves were ignored we need to keep accounting for the score
    if (ignoreLeaf) {
        savedScore = probOfCut * scoreUnseen(depthOfNode, node.getMass()) + (1 - probOfCut) * savedScore;
    }
    if (probOfCut <= 0) {
        pointInsideBox = true;
    } else {
        double newScore = scoreUnseen(depthOfNode, node.getMass());
        for (int i = 0; i < pointToScore.length; i++) {
            double probOfCutInSpikeDirection = differenceInRangeVector[2 * i] / sumOfNewRange;
            directionalAttribution.high[i] = probOfCutInSpikeDirection * newScore + (1 - probOfCut) * directionalAttribution.high[i];
            double probOfCutInDipDirection = differenceInRangeVector[2 * i + 1] / sumOfNewRange;
            directionalAttribution.low[i] = probOfCutInDipDirection * newScore + (1 - probOfCut) * directionalAttribution.low[i];
        }
    }
    if ((hitDuplicates || ignoreLeaf) && (pointInsideBox || depthOfNode == 0)) {
        // final rescaling; this ensures agreement with the ScalarScoreVector
        // the scoreUnseen/scoreSeen should be the same as scoring; other uses need
        // caution.
        directionalAttribution.renormalize(savedScore);
    }
}

Example 8 with IBoundingBoxView

use of com.amazon.randomcutforest.tree.IBoundingBoxView in project random-cut-forest-by-aws by aws.

the class TransductiveScalarScoreVisitor method getWeight.

// for this visitor class the assumption is that the trees are built using the
// same probabilities as are used in scoring. In the application herein
// vecSepBuild
// is the same as vecSepScore as in the accept(node) above; however the function
// is
// written in the more general form so that it can be used for the Simulated
// version as well without any changes.
protected double getWeight(int dim, Function<IBoundingBoxView, double[]> vecSepBuild, final IBoundingBoxView boundingBox) {
    double[] vecSmall = vecSepBuild.apply(boundingBox);
    // the smaller box was built!
    IBoundingBoxView largeBox = boundingBox.getMergedBox(pointToScore);
    double[] vecLarge = vecSepScore.apply(largeBox);
    // the larger box is only scored!
    double sumSmall = 0;
    double sumLarge = 0;
    for (int i = 0; i < pointToScore.length; i++) {
        sumSmall += vecSmall[i];
        sumLarge += vecLarge[i];
    }
    return (boundingBox.getRange(dim) / largeBox.getRange(dim)) * (sumSmall / sumLarge) * (vecLarge[dim] / vecSmall[dim]);
// this can be larger than 1
// For RCFs vecLarge[dim] = largeBox.getRange(dim) and
// vecSmall[dim] = smallBox.getRange(dim)
// sumSmall/sumLarge is the probability of non-separation
}