Predicates and bagging

SMX builds logical predicates from zone scores and uses bagging to stabilize importance rankings across subsamples.

Predicate generation

PredicateGenerator creates two predicates per quantile per zone:

• zone <= threshold

• zone > threshold

from smx import PredicateGenerator

generator = PredicateGenerator(quantiles=[0.25, 0.5, 0.75])
generator.fit(zone_scores)

predicates_df = generator.predicates_df_
indicator_df = generator.indicator_df_

Bagging

PredicateBagger subsamples rows (and optionally predicates) to build bags that feed the metric computations:

from smx import PredicateBagger

bagger = PredicateBagger(n_bags=10, n_samples_fraction=0.8, replace=False, random_seed=42)
bags = bagger.run(zone_scores, y_pred_cal, predicates_df)

Metrics

Two main metrics are provided:

• CovarianceMetric: covariance or mutual information between zone values and predictions

• PerturbationMetric: replace a zone and measure prediction shift

from smx import CovarianceMetric, PerturbationMetric

cov_metric = CovarianceMetric(metric="covariance", threshold=0.01)
rankings = cov_metric.compute(bags)

pert_metric = PerturbationMetric(
    estimator=model,
    Xcalclass_prep=X_cal_prep,
    predicates_df=predicates_df,
    spectral_cuts=spectral_cuts,
    perturbation_mode="median",
    metric="probability_shift",
)
rankings = pert_metric.compute(bags)