Search in sources :

Example 6 with LimeExplainer

use of org.kie.kogito.explainability.local.lime.LimeExplainer in project kogito-apps by kiegroup.

the class FraudScoringDmnLimeExplainerTest method testExplanationStabilityWithOptimization.

@Test
void testExplanationStabilityWithOptimization() throws ExecutionException, InterruptedException, TimeoutException {
    PredictionProvider model = getModel();
    List<PredictionInput> samples = DmnTestUtils.randomFraudScoringInputs();
    List<PredictionOutput> predictionOutputs = model.predictAsync(samples.subList(0, 5)).get();
    List<Prediction> predictions = DataUtils.getPredictions(samples, predictionOutputs);
    long seed = 0;
    LimeConfigOptimizer limeConfigOptimizer = new LimeConfigOptimizer().withDeterministicExecution(true).withStepCountLimit(20);
    Random random = new Random();
    PerturbationContext perturbationContext = new PerturbationContext(seed, random, 1);
    LimeConfig initialConfig = new LimeConfig().withSamples(10).withPerturbationContext(perturbationContext);
    LimeConfig optimizedConfig = limeConfigOptimizer.optimize(initialConfig, predictions, model);
    assertThat(optimizedConfig).isNotSameAs(initialConfig);
    LimeExplainer limeExplainer = new LimeExplainer(optimizedConfig);
    PredictionInput testPredictionInput = getTestInput();
    List<PredictionOutput> testPredictionOutputs = model.predictAsync(List.of(testPredictionInput)).get(Config.INSTANCE.getAsyncTimeout(), Config.INSTANCE.getAsyncTimeUnit());
    Prediction instance = new SimplePrediction(testPredictionInput, testPredictionOutputs.get(0));
    assertDoesNotThrow(() -> ValidationUtils.validateLocalSaliencyStability(model, instance, limeExplainer, 1, 0.5, 0.5));
}
Also used : SimplePrediction(org.kie.kogito.explainability.model.SimplePrediction) PerturbationContext(org.kie.kogito.explainability.model.PerturbationContext) PredictionInput(org.kie.kogito.explainability.model.PredictionInput) LimeExplainer(org.kie.kogito.explainability.local.lime.LimeExplainer) Prediction(org.kie.kogito.explainability.model.Prediction) SimplePrediction(org.kie.kogito.explainability.model.SimplePrediction) PredictionProvider(org.kie.kogito.explainability.model.PredictionProvider) LimeConfig(org.kie.kogito.explainability.local.lime.LimeConfig) Random(java.util.Random) PredictionOutput(org.kie.kogito.explainability.model.PredictionOutput) LimeConfigOptimizer(org.kie.kogito.explainability.local.lime.optim.LimeConfigOptimizer) Test(org.junit.jupiter.api.Test)

Example 7 with LimeExplainer

use of org.kie.kogito.explainability.local.lime.LimeExplainer in project kogito-apps by kiegroup.

the class FraudScoringDmnLimeExplainerTest method testExplanationWeightedStabilityWithOptimization.

@Test
void testExplanationWeightedStabilityWithOptimization() throws ExecutionException, InterruptedException, TimeoutException {
    PredictionProvider model = getModel();
    List<PredictionInput> samples = DmnTestUtils.randomFraudScoringInputs();
    List<PredictionOutput> predictionOutputs = model.predictAsync(samples.subList(0, 5)).get();
    List<Prediction> predictions = DataUtils.getPredictions(samples, predictionOutputs);
    long seed = 0;
    LimeConfigOptimizer limeConfigOptimizer = new LimeConfigOptimizer().withDeterministicExecution(true).withWeightedStability(0.4, 0.6).withStepCountLimit(10);
    Random random = new Random();
    PerturbationContext perturbationContext = new PerturbationContext(seed, random, 1);
    LimeConfig initialConfig = new LimeConfig().withSamples(10).withPerturbationContext(perturbationContext);
    LimeConfig optimizedConfig = limeConfigOptimizer.optimize(initialConfig, predictions, model);
    assertThat(optimizedConfig).isNotSameAs(initialConfig);
    LimeExplainer limeExplainer = new LimeExplainer(optimizedConfig);
    PredictionInput testPredictionInput = getTestInput();
    List<PredictionOutput> testPredictionOutputs = model.predictAsync(List.of(testPredictionInput)).get(Config.INSTANCE.getAsyncTimeout(), Config.INSTANCE.getAsyncTimeUnit());
    Prediction instance = new SimplePrediction(testPredictionInput, testPredictionOutputs.get(0));
    assertDoesNotThrow(() -> ValidationUtils.validateLocalSaliencyStability(model, instance, limeExplainer, 1, 0.6, 0.4));
}
Also used : SimplePrediction(org.kie.kogito.explainability.model.SimplePrediction) PerturbationContext(org.kie.kogito.explainability.model.PerturbationContext) PredictionInput(org.kie.kogito.explainability.model.PredictionInput) LimeExplainer(org.kie.kogito.explainability.local.lime.LimeExplainer) Prediction(org.kie.kogito.explainability.model.Prediction) SimplePrediction(org.kie.kogito.explainability.model.SimplePrediction) PredictionProvider(org.kie.kogito.explainability.model.PredictionProvider) LimeConfig(org.kie.kogito.explainability.local.lime.LimeConfig) Random(java.util.Random) PredictionOutput(org.kie.kogito.explainability.model.PredictionOutput) LimeConfigOptimizer(org.kie.kogito.explainability.local.lime.optim.LimeConfigOptimizer) Test(org.junit.jupiter.api.Test)

Example 8 with LimeExplainer

use of org.kie.kogito.explainability.local.lime.LimeExplainer in project kogito-apps by kiegroup.

the class FraudScoringDmnLimeExplainerTest method testFraudScoringDMNExplanation.

@Test
void testFraudScoringDMNExplanation() throws ExecutionException, InterruptedException, TimeoutException {
    PredictionProvider model = getModel();
    PredictionInput predictionInput = getTestInput();
    List<PredictionOutput> predictionOutputs = model.predictAsync(List.of(predictionInput)).get(Config.INSTANCE.getAsyncTimeout(), Config.INSTANCE.getAsyncTimeUnit());
    Prediction prediction = new SimplePrediction(predictionInput, predictionOutputs.get(0));
    Random random = new Random();
    PerturbationContext perturbationContext = new PerturbationContext(0L, random, 1);
    LimeConfig limeConfig = new LimeConfig().withSamples(10).withPerturbationContext(perturbationContext);
    LimeExplainer limeExplainer = new LimeExplainer(limeConfig);
    Map<String, Saliency> saliencyMap = limeExplainer.explainAsync(prediction, model).get(Config.INSTANCE.getAsyncTimeout(), Config.INSTANCE.getAsyncTimeUnit());
    for (Saliency saliency : saliencyMap.values()) {
        assertNotNull(saliency);
        List<FeatureImportance> topFeatures = saliency.getTopFeatures(4);
        double topScore = Math.abs(topFeatures.stream().map(FeatureImportance::getScore).findFirst().orElse(0d));
        if (!topFeatures.isEmpty() && topScore > 0) {
            double v = ExplainabilityMetrics.impactScore(model, prediction, topFeatures);
            // checks the drop of important features triggers a flipped prediction (or a significant drop in the output score).
            assertThat(v).isPositive();
        }
    }
    assertDoesNotThrow(() -> ValidationUtils.validateLocalSaliencyStability(model, prediction, limeExplainer, 1, 0.4, // set to 0.4 since "Last Transaction" is inherently unstable output
    0.4));
    String decision = "Risk Score";
    List<PredictionInput> inputs = new ArrayList<>();
    for (int n = 0; n < 10; n++) {
        inputs.add(new PredictionInput(DataUtils.perturbFeatures(predictionInput.getFeatures(), perturbationContext)));
    }
    DataDistribution distribution = new PredictionInputsDataDistribution(inputs);
    int k = 2;
    int chunkSize = 2;
    double f1 = ExplainabilityMetrics.getLocalSaliencyF1(decision, model, limeExplainer, distribution, k, chunkSize);
    AssertionsForClassTypes.assertThat(f1).isBetween(0.5d, 1d);
}
Also used : SimplePrediction(org.kie.kogito.explainability.model.SimplePrediction) PerturbationContext(org.kie.kogito.explainability.model.PerturbationContext) PredictionInput(org.kie.kogito.explainability.model.PredictionInput) LimeExplainer(org.kie.kogito.explainability.local.lime.LimeExplainer) Prediction(org.kie.kogito.explainability.model.Prediction) SimplePrediction(org.kie.kogito.explainability.model.SimplePrediction) ArrayList(java.util.ArrayList) Saliency(org.kie.kogito.explainability.model.Saliency) PredictionProvider(org.kie.kogito.explainability.model.PredictionProvider) LimeConfig(org.kie.kogito.explainability.local.lime.LimeConfig) Random(java.util.Random) FeatureImportance(org.kie.kogito.explainability.model.FeatureImportance) PredictionInputsDataDistribution(org.kie.kogito.explainability.model.PredictionInputsDataDistribution) DataDistribution(org.kie.kogito.explainability.model.DataDistribution) PredictionOutput(org.kie.kogito.explainability.model.PredictionOutput) PredictionInputsDataDistribution(org.kie.kogito.explainability.model.PredictionInputsDataDistribution) Test(org.junit.jupiter.api.Test)

Example 9 with LimeExplainer

use of org.kie.kogito.explainability.local.lime.LimeExplainer in project kogito-apps by kiegroup.

the class LoanEligibilityDmnLimeExplainerTest method testLoanEligibilityDMNExplanation.

@Test
void testLoanEligibilityDMNExplanation() throws ExecutionException, InterruptedException, TimeoutException {
    PredictionProvider model = getModel();
    PredictionInput predictionInput = getTestInput();
    List<PredictionOutput> predictionOutputs = model.predictAsync(List.of(predictionInput)).get(Config.INSTANCE.getAsyncTimeout(), Config.INSTANCE.getAsyncTimeUnit());
    Prediction prediction = new SimplePrediction(predictionInput, predictionOutputs.get(0));
    Random random = new Random();
    PerturbationContext perturbationContext = new PerturbationContext(0L, random, 1);
    LimeConfig limeConfig = new LimeConfig().withPerturbationContext(perturbationContext);
    LimeExplainer limeExplainer = new LimeExplainer(limeConfig);
    Map<String, Saliency> saliencyMap = limeExplainer.explainAsync(prediction, model).get(Config.INSTANCE.getAsyncTimeout(), Config.INSTANCE.getAsyncTimeUnit());
    for (Saliency saliency : saliencyMap.values()) {
        assertNotNull(saliency);
    }
    assertDoesNotThrow(() -> ValidationUtils.validateLocalSaliencyStability(model, prediction, limeExplainer, 1, 0.4, 0.4));
    String decision = "Eligibility";
    List<PredictionInput> inputs = new ArrayList<>();
    for (int n = 0; n < 10; n++) {
        inputs.add(new PredictionInput(DataUtils.perturbFeatures(predictionInput.getFeatures(), perturbationContext)));
    }
    DataDistribution distribution = new PredictionInputsDataDistribution(inputs);
    int k = 2;
    int chunkSize = 2;
    double f1 = ExplainabilityMetrics.getLocalSaliencyF1(decision, model, limeExplainer, distribution, k, chunkSize);
    AssertionsForClassTypes.assertThat(f1).isBetween(0.5d, 1d);
}
Also used : SimplePrediction(org.kie.kogito.explainability.model.SimplePrediction) PerturbationContext(org.kie.kogito.explainability.model.PerturbationContext) PredictionInput(org.kie.kogito.explainability.model.PredictionInput) LimeExplainer(org.kie.kogito.explainability.local.lime.LimeExplainer) Prediction(org.kie.kogito.explainability.model.Prediction) SimplePrediction(org.kie.kogito.explainability.model.SimplePrediction) ArrayList(java.util.ArrayList) Saliency(org.kie.kogito.explainability.model.Saliency) PredictionProvider(org.kie.kogito.explainability.model.PredictionProvider) LimeConfig(org.kie.kogito.explainability.local.lime.LimeConfig) Random(java.util.Random) PredictionInputsDataDistribution(org.kie.kogito.explainability.model.PredictionInputsDataDistribution) DataDistribution(org.kie.kogito.explainability.model.DataDistribution) PredictionOutput(org.kie.kogito.explainability.model.PredictionOutput) PredictionInputsDataDistribution(org.kie.kogito.explainability.model.PredictionInputsDataDistribution) Test(org.junit.jupiter.api.Test)

Example 10 with LimeExplainer

use of org.kie.kogito.explainability.local.lime.LimeExplainer in project kogito-apps by kiegroup.

the class LoanEligibilityDmnLimeExplainerTest method testExplanationWeightedStabilityWithOptimization.

@Test
void testExplanationWeightedStabilityWithOptimization() throws ExecutionException, InterruptedException, TimeoutException {
    PredictionProvider model = getModel();
    List<PredictionInput> samples = DmnTestUtils.randomLoanEligibilityInputs();
    List<PredictionOutput> predictionOutputs = model.predictAsync(samples.subList(0, 5)).get();
    List<Prediction> predictions = DataUtils.getPredictions(samples, predictionOutputs);
    long seed = 0;
    LimeConfigOptimizer limeConfigOptimizer = new LimeConfigOptimizer().withDeterministicExecution(true).withWeightedStability(0.4, 0.6).withStepCountLimit(10);
    Random random = new Random();
    PerturbationContext perturbationContext = new PerturbationContext(seed, random, 1);
    LimeConfig initialConfig = new LimeConfig().withPerturbationContext(perturbationContext);
    LimeConfig optimizedConfig = limeConfigOptimizer.optimize(initialConfig, predictions, model);
    assertThat(optimizedConfig).isNotSameAs(initialConfig);
    LimeExplainer limeExplainer = new LimeExplainer(optimizedConfig);
    PredictionInput testPredictionInput = getTestInput();
    List<PredictionOutput> testPredictionOutputs = model.predictAsync(List.of(testPredictionInput)).get(Config.INSTANCE.getAsyncTimeout(), Config.INSTANCE.getAsyncTimeUnit());
    Prediction instance = new SimplePrediction(testPredictionInput, testPredictionOutputs.get(0));
    assertDoesNotThrow(() -> ValidationUtils.validateLocalSaliencyStability(model, instance, limeExplainer, 1, 0.6, 0.4));
}
Also used : SimplePrediction(org.kie.kogito.explainability.model.SimplePrediction) PerturbationContext(org.kie.kogito.explainability.model.PerturbationContext) PredictionInput(org.kie.kogito.explainability.model.PredictionInput) LimeExplainer(org.kie.kogito.explainability.local.lime.LimeExplainer) Prediction(org.kie.kogito.explainability.model.Prediction) SimplePrediction(org.kie.kogito.explainability.model.SimplePrediction) PredictionProvider(org.kie.kogito.explainability.model.PredictionProvider) LimeConfig(org.kie.kogito.explainability.local.lime.LimeConfig) Random(java.util.Random) PredictionOutput(org.kie.kogito.explainability.model.PredictionOutput) LimeConfigOptimizer(org.kie.kogito.explainability.local.lime.optim.LimeConfigOptimizer) Test(org.junit.jupiter.api.Test)

Aggregations

LimeExplainer (org.kie.kogito.explainability.local.lime.LimeExplainer)42 Prediction (org.kie.kogito.explainability.model.Prediction)38 Test (org.junit.jupiter.api.Test)37 LimeConfig (org.kie.kogito.explainability.local.lime.LimeConfig)36 PredictionProvider (org.kie.kogito.explainability.model.PredictionProvider)36 SimplePrediction (org.kie.kogito.explainability.model.SimplePrediction)35 PredictionInput (org.kie.kogito.explainability.model.PredictionInput)34 Random (java.util.Random)33 PerturbationContext (org.kie.kogito.explainability.model.PerturbationContext)33 PredictionOutput (org.kie.kogito.explainability.model.PredictionOutput)33 LimeConfigOptimizer (org.kie.kogito.explainability.local.lime.optim.LimeConfigOptimizer)19 Saliency (org.kie.kogito.explainability.model.Saliency)16 PredictionInputsDataDistribution (org.kie.kogito.explainability.model.PredictionInputsDataDistribution)13 DataDistribution (org.kie.kogito.explainability.model.DataDistribution)12 ArrayList (java.util.ArrayList)9 Feature (org.kie.kogito.explainability.model.Feature)7 ExecutionException (java.util.concurrent.ExecutionException)5 TimeoutException (java.util.concurrent.TimeoutException)5 FeatureImportance (org.kie.kogito.explainability.model.FeatureImportance)5 InputStreamReader (java.io.InputStreamReader)4