Examples with FeatureDistribution org.kie.kogito.explainability.model.FeatureDistribution used on opensource projects

Example 11 with FeatureDistribution

use of org.kie.kogito.explainability.model.FeatureDistribution in project kogito-apps by kiegroup.

the class DataUtilsTest method testBootstrap.

@Test
void testBootstrap() {
    List<Value> values = new ArrayList<>();
    PerturbationContext perturbationContext = new PerturbationContext(random, 1);
    for (int i = 0; i < 4; i++) {
        values.add(Type.NUMBER.randomValue(perturbationContext));
    }
    Feature mockedNumericFeature = TestUtils.getMockedNumericFeature();
    DataDistribution dataDistribution = new IndependentFeaturesDataDistribution(List.of(new GenericFeatureDistribution(mockedNumericFeature, values)));
    Map<String, FeatureDistribution> featureDistributionMap = DataUtils.boostrapFeatureDistributions(dataDistribution, perturbationContext, 10, 1, 500, new HashMap<>());
    assertThat(featureDistributionMap).isNotNull();
    assertThat(featureDistributionMap).isNotEmpty();
    FeatureDistribution actual = featureDistributionMap.get(mockedNumericFeature.getName());
    assertThat(actual).isNotNull();
    List<Value> allSamples = actual.getAllSamples();
    assertThat(allSamples).isNotNull();
    assertThat(allSamples).hasSize(10);
}

Example 12 with FeatureDistribution

use of org.kie.kogito.explainability.model.FeatureDistribution in project kogito-apps by kiegroup.

the class DataUtilsTest method testRandomDistributionGeneration.

@Test
void testRandomDistributionGeneration() {
    DataDistribution dataDistribution = DataUtils.generateRandomDataDistribution(10, 10, random);
    assertNotNull(dataDistribution);
    assertNotNull(dataDistribution.asFeatureDistributions());
    for (FeatureDistribution featureDistribution : dataDistribution.asFeatureDistributions()) {
        assertNotNull(featureDistribution);
    }
}

Example 13 with FeatureDistribution

use of org.kie.kogito.explainability.model.FeatureDistribution in project kogito-apps by kiegroup.

the class DataUtils method generateRandomDataDistribution.

/**
 * Generate a random data distribution.
 *
 * @param noOfFeatures number of features
 * @param distributionSize number of samples for each feature
 * @return a data distribution
 */
public static DataDistribution generateRandomDataDistribution(int noOfFeatures, int distributionSize, Random random) {
    List<FeatureDistribution> featureDistributions = new LinkedList<>();
    for (int i = 0; i < noOfFeatures; i++) {
        double[] doubles = generateData(random.nextDouble(), random.nextDouble(), distributionSize, random);
        Feature feature = FeatureFactory.newNumericalFeature("f_" + i, Double.NaN);
        FeatureDistribution featureDistribution = new NumericFeatureDistribution(feature, doubles);
        featureDistributions.add(featureDistribution);
    }
    return new IndependentFeaturesDataDistribution(featureDistributions);
}

Example 14 with FeatureDistribution

use of org.kie.kogito.explainability.model.FeatureDistribution in project kogito-apps by kiegroup.

the class PartialDependencePlotExplainer method explainFromDataDistribution.

private List<PartialDependenceGraph> explainFromDataDistribution(PredictionProvider model, int outputSize, DataDistribution dataDistribution) throws InterruptedException, ExecutionException, TimeoutException {
    long start = System.currentTimeMillis();
    List<PartialDependenceGraph> pdps = new ArrayList<>();
    List<FeatureDistribution> featureDistributions = dataDistribution.asFeatureDistributions();
    // fetch entire data distributions for all features
    List<PredictionInput> trainingData = dataDistribution.sample(config.getSeriesLength());
    // create a PDP for each feature
    for (FeatureDistribution featureDistribution : featureDistributions) {
        // generate (further) samples for the feature under analysis
        // TBD: maybe just reuse trainingData
        List<Value> xsValues = featureDistribution.sample(config.getSeriesLength()).stream().sorted(// sort alphanumerically (if Value#asNumber is NaN)
        Comparator.comparing(Value::asString)).sorted(// sort by natural order
        (v1, v2) -> Comparator.comparingDouble(Value::asNumber).compare(v1, v2)).distinct().collect(Collectors.toList());
        List<Feature> featureXSvalues = // transform sampled Values into Features
        xsValues.stream().map(v -> FeatureFactory.copyOf(featureDistribution.getFeature(), v)).collect(Collectors.toList());
        // create a PDP for each feature and each output
        for (int outputIndex = 0; outputIndex < outputSize; outputIndex++) {
            PartialDependenceGraph partialDependenceGraph = getPartialDependenceGraph(model, trainingData, xsValues, featureXSvalues, outputIndex);
            pdps.add(partialDependenceGraph);
        }
    }
    long end = System.currentTimeMillis();
    LOGGER.debug("explanation time: {}ms", (end - start));
    return pdps;
}

Example 15 with FeatureDistribution

use of org.kie.kogito.explainability.model.FeatureDistribution in project kogito-apps by kiegroup.

the class LimeExplainer method getPerturbedInputs.

private List<PredictionInput> getPerturbedInputs(List<Feature> features, LimeConfig executionConfig, PredictionProvider predictionProvider) {
    List<PredictionInput> perturbedInputs = new ArrayList<>();
    int size = executionConfig.getNoOfSamples();
    DataDistribution dataDistribution = executionConfig.getDataDistribution();
    Map<String, FeatureDistribution> featureDistributionsMap;
    PerturbationContext perturbationContext = executionConfig.getPerturbationContext();
    if (!dataDistribution.isEmpty()) {
        Map<String, HighScoreNumericFeatureZones> numericFeatureZonesMap;
        int max = executionConfig.getBoostrapInputs();
        if (executionConfig.isHighScoreFeatureZones()) {
            numericFeatureZonesMap = HighScoreNumericFeatureZonesProvider.getHighScoreFeatureZones(dataDistribution, predictionProvider, features, max);
        } else {
            numericFeatureZonesMap = new HashMap<>();
        }
        // generate feature distributions, if possible
        featureDistributionsMap = DataUtils.boostrapFeatureDistributions(dataDistribution, perturbationContext, 2 * size, 1, Math.min(size, max), numericFeatureZonesMap);
    } else {
        featureDistributionsMap = new HashMap<>();
    }
    for (int i = 0; i < size; i++) {
        List<Feature> newFeatures = DataUtils.perturbFeatures(features, perturbationContext, featureDistributionsMap);
        perturbedInputs.add(new PredictionInput(newFeatures));
    }
    return perturbedInputs;
}