Function CalcSageValues

catboost/libs/fstr/sage_values.cpp:343–427

}

TVector<TVector<double>> CalcSageValues(
    const TFullModel& model,
    const TDataProvider& dataset,
    int logPeriod,
    NPar::ILocalExecutor* localExecutor,
    size_t nSamples,
    size_t batchSize,
    bool detectConvergence)
{
    CB_ENSURE(model.ModelTrees->GetDimensionsCount() == 1, "Model must not be trained for multiclassification");

    // setting algorithm params
    TRestorableFastRng64 rand(228);
    size_t featuresCount = dataset.MetaInfo.GetFeatureCount();
    batchSize = Min(batchSize, size_t(dataset.GetObjectCount()));
    auto featuresLayout = dataset.MetaInfo.FeaturesLayout;
    const double convergenceThreshold = 0.1;

    FullSubsetIndexingPtrWrapper fullSubsetIndexingPtrWrapper;
    MarginalImputer imputer(dataset, localExecutor, &rand);

    // creating loss holder
    NCatboostOptions::TLossDescription metricDescription;
    NCatboostOptions::TLossDescription lossDescription;
    bool needYetiRankPairs = false;
    THolder<IMetric> metric;

    CreateMetricAndLossDescriptionForLossChange(
        model,
        &metricDescription,
        &lossDescription,
        &needYetiRankPairs,
        &metric
    );

    CB_ENSURE_INTERNAL(metric->IsAdditiveMetric(), "Loss function must be additive");

    TVector<THolder<IMetric>> metrics;
    metrics.push_back(std::move(metric));

    // main algorithm: calculating sage values
    TImportanceLogger samplingIterationsLogger(nSamples, "sampling iterations passed",
                                               "Calculating SAGE values...", logPeriod);
    TProfileInfo samplingItertionsProfile(nSamples);
    TVector<TVector<double>> sageValues(featuresCount, TVector<double>{0});
    for (size_t i = 0; i < nSamples; ++i) {
        samplingItertionsProfile.StartIterationBlock();

        // sampling batch of dataset elements
        auto datasetBatch = GetRandomDatasetBatch(dataset, batchSize, &rand, localExecutor);

        // generating features permutation
        auto featuresPermutation = GenerateFeaturesPermutation(featuresLayout, &rand);

        // running approximation algorithm
        double previousLoss = CalculateModelLoss(model, datasetBatch, metrics, &rand, localExecutor);
        for (size_t j = 0; j < featuresCount; ++j) {
            // preparing dataset, sampling disabled features