use of org.kie.kogito.explainability.model.Output in project kogito-apps by kiegroup.
the class CounterfactualScoreCalculatorTest method DoubleDistanceDifferentValueThreshold.
@ParameterizedTest
@ValueSource(ints = { 0, 1, 2, 3, 4 })
void DoubleDistanceDifferentValueThreshold(int seed) {
final Random random = new Random(seed);
double value = random.nextDouble() * 100.0;
Feature x = FeatureFactory.newNumericalFeature("x", value);
Feature y = FeatureFactory.newNumericalFeature("y", value + 100.0);
Output ox = outputFromFeature(x);
Output oy = outputFromFeature(y);
double distance = CounterFactualScoreCalculator.outputDistance(ox, oy, 0.25);
assertEquals(Type.NUMBER, ox.getType());
assertEquals(Type.NUMBER, oy.getType());
assertTrue(distance * distance > 0);
y = FeatureFactory.newNumericalFeature("y", value - 100);
oy = outputFromFeature(y);
distance = CounterFactualScoreCalculator.outputDistance(ox, oy, 0.25);
assertTrue(distance * distance > 0);
}
use of org.kie.kogito.explainability.model.Output in project kogito-apps by kiegroup.
the class CounterfactualScoreCalculatorTest method binaryDistanceSameValue.
@ParameterizedTest
@ValueSource(ints = { 0, 1, 2, 3, 4 })
void binaryDistanceSameValue(int seed) {
final Random random = new Random(seed);
final ByteBuffer value = ByteBuffer.wrap("foo".getBytes());
Feature x = FeatureFactory.newBinaryFeature("x", value);
Feature y = FeatureFactory.newBinaryFeature("y", value);
Output ox = outputFromFeature(x);
Output oy = outputFromFeature(y);
double distance = CounterFactualScoreCalculator.outputDistance(ox, oy);
assertEquals(Type.BINARY, ox.getType());
assertEquals(0.0, distance);
// Use a random threshold, mustn't make a difference
distance = CounterFactualScoreCalculator.outputDistance(ox, oy, random.nextDouble());
assertEquals(0.0, distance);
}
use of org.kie.kogito.explainability.model.Output in project kogito-apps by kiegroup.
the class ExplainabilityMetrics method getLocalSaliencyRecall.
/**
* Evaluate the recall of a local saliency explainer on a given model.
* Get the predictions having outputs with the highest score for the given decision and pair them with predictions
* whose outputs have the lowest score for the same decision.
* Get the top k (most important) features (according to the saliency) for the most important outputs and
* "paste" them on each paired input corresponding to an output with low score (for the target decision).
* Perform prediction on the "masked" input, if the output on the masked input is equals to the output for the
* input the mask features were take from, that's considered a true positive, otherwise it's a false positive.
* see Section 3.2.1 of https://openreview.net/attachment?id=B1xBAA4FwH&name=original_pdf
*
* @param outputName decision to evaluate recall for
* @param predictionProvider the prediction provider to test
* @param localExplainer the explainer to evaluate
* @param dataDistribution the data distribution used to obtain inputs for evaluation
* @param k the no. of features to extract
* @param chunkSize the size of the chunk of predictions to use for evaluation
* @return the saliency recall
*/
public static double getLocalSaliencyRecall(String outputName, PredictionProvider predictionProvider, LocalExplainer<Map<String, Saliency>> localExplainer, DataDistribution dataDistribution, int k, int chunkSize) throws InterruptedException, ExecutionException, TimeoutException {
// get all samples from the data distribution
List<Prediction> sorted = DataUtils.getScoreSortedPredictions(outputName, predictionProvider, dataDistribution);
// get the top and bottom 'chunkSize' predictions
List<Prediction> topChunk = new ArrayList<>(sorted.subList(0, chunkSize));
List<Prediction> bottomChunk = new ArrayList<>(sorted.subList(sorted.size() - chunkSize, sorted.size()));
double truePositives = 0;
double falseNegatives = 0;
int currentChunk = 0;
// input, then feed the model with this masked input and check the output is equals to the top scored one.
for (Prediction prediction : topChunk) {
Optional<Output> optionalOutput = prediction.getOutput().getByName(outputName);
if (optionalOutput.isPresent()) {
Output output = optionalOutput.get();
Map<String, Saliency> stringSaliencyMap = localExplainer.explainAsync(prediction, predictionProvider).get(Config.DEFAULT_ASYNC_TIMEOUT, Config.DEFAULT_ASYNC_TIMEUNIT);
if (stringSaliencyMap.containsKey(outputName)) {
Saliency saliency = stringSaliencyMap.get(outputName);
List<FeatureImportance> topFeatures = saliency.getPerFeatureImportance().stream().sorted((f1, f2) -> Double.compare(f2.getScore(), f1.getScore())).limit(k).collect(Collectors.toList());
PredictionInput input = bottomChunk.get(currentChunk).getInput();
PredictionInput maskedInput = maskInput(topFeatures, input);
List<PredictionOutput> predictionOutputList = predictionProvider.predictAsync(List.of(maskedInput)).get(Config.DEFAULT_ASYNC_TIMEOUT, Config.DEFAULT_ASYNC_TIMEUNIT);
if (!predictionOutputList.isEmpty()) {
PredictionOutput predictionOutput = predictionOutputList.get(0);
Optional<Output> optionalNewOutput = predictionOutput.getByName(outputName);
if (optionalNewOutput.isPresent()) {
Output newOutput = optionalOutput.get();
if (output.getValue().equals(newOutput.getValue())) {
truePositives++;
} else {
falseNegatives++;
}
}
}
currentChunk++;
}
}
}
if ((truePositives + falseNegatives) > 0) {
return truePositives / (truePositives + falseNegatives);
} else {
// if topChunk is empty or the target output (by name) is not an output of the model.
return Double.NaN;
}
}
use of org.kie.kogito.explainability.model.Output in project kogito-apps by kiegroup.
the class ExplainabilityMetrics method getLocalSaliencyPrecision.
/**
* Evaluate the precision of a local saliency explainer on a given model.
* Get the predictions having outputs with the lowest score for the given decision and pair them with predictions
* whose outputs have the highest score for the same decision.
* Get the bottom k (less important) features (according to the saliency) for the less important outputs and
* "paste" them on each paired input corresponding to an output with high score (for the target decision).
* Perform prediction on the "masked" input, if the output changes that's considered a false negative, otherwise
* it's a true positive.
* see Section 3.2.1 of https://openreview.net/attachment?id=B1xBAA4FwH&name=original_pdf
*
* @param outputName decision to evaluate recall for
* @param predictionProvider the prediction provider to test
* @param localExplainer the explainer to evaluate
* @param dataDistribution the data distribution used to obtain inputs for evaluation
* @param k the no. of features to extract
* @param chunkSize the size of the chunk of predictions to use for evaluation
* @return the saliency precision
*/
public static double getLocalSaliencyPrecision(String outputName, PredictionProvider predictionProvider, LocalExplainer<Map<String, Saliency>> localExplainer, DataDistribution dataDistribution, int k, int chunkSize) throws InterruptedException, ExecutionException, TimeoutException {
List<Prediction> sorted = DataUtils.getScoreSortedPredictions(outputName, predictionProvider, dataDistribution);
// get the top and bottom 'chunkSize' predictions
List<Prediction> topChunk = new ArrayList<>(sorted.subList(0, chunkSize));
List<Prediction> bottomChunk = new ArrayList<>(sorted.subList(sorted.size() - chunkSize, sorted.size()));
double truePositives = 0;
double falsePositives = 0;
int currentChunk = 0;
for (Prediction prediction : bottomChunk) {
Map<String, Saliency> stringSaliencyMap = localExplainer.explainAsync(prediction, predictionProvider).get(Config.DEFAULT_ASYNC_TIMEOUT, Config.DEFAULT_ASYNC_TIMEUNIT);
if (stringSaliencyMap.containsKey(outputName)) {
Saliency saliency = stringSaliencyMap.get(outputName);
List<FeatureImportance> topFeatures = saliency.getPerFeatureImportance().stream().sorted(Comparator.comparingDouble(FeatureImportance::getScore)).limit(k).collect(Collectors.toList());
Prediction topPrediction = topChunk.get(currentChunk);
PredictionInput input = topPrediction.getInput();
PredictionInput maskedInput = maskInput(topFeatures, input);
List<PredictionOutput> predictionOutputList = predictionProvider.predictAsync(List.of(maskedInput)).get(Config.DEFAULT_ASYNC_TIMEOUT, Config.DEFAULT_ASYNC_TIMEUNIT);
if (!predictionOutputList.isEmpty()) {
PredictionOutput predictionOutput = predictionOutputList.get(0);
Optional<Output> newOptionalOutput = predictionOutput.getByName(outputName);
if (newOptionalOutput.isPresent()) {
Output newOutput = newOptionalOutput.get();
Optional<Output> optionalOutput = topPrediction.getOutput().getByName(outputName);
if (optionalOutput.isPresent()) {
Output output = optionalOutput.get();
if (output.getValue().equals(newOutput.getValue())) {
truePositives++;
} else {
falsePositives++;
}
}
}
}
currentChunk++;
}
}
if ((truePositives + falsePositives) > 0) {
return truePositives / (truePositives + falsePositives);
} else {
// if bottomChunk is empty or the target output (by name) is not an output of the model.
return Double.NaN;
}
}
use of org.kie.kogito.explainability.model.Output in project kogito-apps by kiegroup.
the class FairnessMetrics method individualConsistency.
/**
* Calculate individual fairness in terms of consistency of predictions across similar inputs.
*
* @param proximityFunction a function that finds the top k similar inputs, given a reference input and a list of inputs
* @param samples a list of inputs to be tested for consistency
* @param predictionProvider the model under inspection
* @return the consistency measure
* @throws ExecutionException if any error occurs during model prediction
* @throws InterruptedException if timeout or other interruption issues occur during model prediction
*/
public static double individualConsistency(BiFunction<PredictionInput, List<PredictionInput>, List<PredictionInput>> proximityFunction, List<PredictionInput> samples, PredictionProvider predictionProvider) throws ExecutionException, InterruptedException {
double consistency = 1;
for (PredictionInput input : samples) {
List<PredictionOutput> predictionOutputs = predictionProvider.predictAsync(List.of(input)).get();
PredictionOutput predictionOutput = predictionOutputs.get(0);
List<PredictionInput> neighbors = proximityFunction.apply(input, samples);
List<PredictionOutput> neighborsOutputs = predictionProvider.predictAsync(neighbors).get();
for (Output output : predictionOutput.getOutputs()) {
Value originalValue = output.getValue();
for (PredictionOutput neighborOutput : neighborsOutputs) {
Output currentOutput = neighborOutput.getByName(output.getName()).orElse(null);
if (currentOutput != null && !originalValue.equals(currentOutput.getValue())) {
consistency -= 1f / (neighbors.size() * predictionOutput.getOutputs().size() * samples.size());
}
}
}
}
return consistency;
}
Aggregations