Task-First Method Selection Matrix¶

← Guide Overview | Choosing by Constraint →

This guide is the user-facing version of the adoption audit capability matrix.

Design intent:

Start from the problem you have, not a modeling ideology
Keep SWAG, BNN, and MDN as peer methods in the catalog
Use evidence-based maturity labels (tests/docs/examples), not family-based labels
Keep Bayesian concepts optional for onboarding

How To Use This Page¶

Find your primary problem in the task matrix.
Start with the recommended defaults.
Use the method matrix to shortlist alternatives when you need:
multimodal outputs
explicit decomposition (epistemic vs aleatoric)
stronger tail/imbalance handling
calibrated intervals / conformal guarantees
OOD scoring

For constructor signatures, scoring rules, and parameter definitions, cross-check symbols against the Losses API and Metrics API. Generated tables below use plain symbol names for readability.

import torchregress as tr

# Start from the task/problem you have:
# - robust regression
# - calibration / conformal intervals
# - multimodal targets (MDN / flows)
# - OOD / selective prediction
#
# Then choose a method family and validate with metrics/examples.
loss = tr.losses.WeightedHuberLoss()

Task Matrix (Recommended Starting Points)¶

Generated provenance: tools/render_method_catalog.py:render_task_matrix_table Source artifacts: reports/method_catalog_latest.json Generated date: 2026-11-08

Task / Constraint	Recommended Start	Strong Alternatives	Notes
Clean regression baseline	`WeightedMSELoss`	`WeightedHuberLoss`	Start simple; add UQ only if needed.
Outliers / robust regression	`WeightedHuberLoss`	`CauchyLoss`, `TukeyBiweightLoss`, `CharbonnierLoss`	Huber is the best default tradeoff.
Worst-case / tail-focused robust regression	`CVaRLoss`	`WeightedHuberLoss`, `CauchyLoss`, `TukeyBiweightLoss`	CVaR optimises the average of the worst α fraction of per-sample losses, directly shrinking the upper tail of the error distribution. Start with CVaR(base_loss='mse', alpha=0.1) and tune alpha for the desired tail-robustness tradeoff. For multi-output targets, CVaR selects the worst samples (not per-element errors), preserving sample-wise semantics.
Unknown noise regime / learnable loss shape	`AdaptiveRobustLoss`	`BarronLoss`, `CVaRLoss`, `WeightedHuberLoss`	AdaptiveRobustLoss jointly optimises the Barron shape parameter α and scale alongside the model weights so the penalty function adapts to the observed noise. Add `loss_fn.parameters()` to the optimizer. Useful as a diagnostic (what α does the data prefer?) or when the noise regime is genuinely unknown a priori. For a fixed α sweep instead, use BarronLoss.
Heteroscedastic noise (aleatoric UQ)	`GaussianCRPSLoss`	`GaussianNLLLoss`, `HeteroscedasticEnsembleModel`, `MixtureDensityLoss`	Astronomical benchmarks favor CRPS-trained Gaussian heads as the safest calibrated Gaussian baseline.
Epistemic uncertainty	`BaseEnsembleModel`	`HeteroscedasticBatchEnsembleModel`, `BinnedPDFEnsembleModel`, `MDNEnsembleModel`, `SWAG`, `BayesianNeuralNetwork`, `MCDropoutWrapper`	Ensembles are easiest operationally.
Low-shot / streaming linear head on fixed features	`BayesianLinearHead`	`RecursiveBayesianHead`, `WeightedMSELoss (ridge MAP, matched L2)`	Conjugate exact BLR for last-layer adaptation; synthetic RMSE/NLL and drift sweeps live under examples/benchmarks/. Prefer ensembles/SWAG/BNN when epistemic UQ must track representation-level ambiguity.
Epistemic + aleatoric decomposition	`HeteroscedasticEnsembleModel`	`HeteroscedasticBNN`, `MDNEnsembleModel`, `HeteroscedasticBatchEnsembleModel`	Requires both model-disagreement and per-member variance/distribution modeling.
Multimodal targets	`MixtureDensityLoss`	`MDNEnsembleModel`, `BinnedPDFEnsembleModel`, `NormalizingFlowLoss`	MDN is usually easier to debug first; ensembles of MDN or ordered-bin heads are the next move when mode averaging matters.
Non-Gaussian / skewed tails	`QuantileLoss` / `ExpectileLoss` / `TweedieLoss`	`MixtureDensityLoss`, `NormalizingFlowLoss`	Choose by target support and evaluation metric.
Multi-target correlated outputs	`MultivariateGaussianLoss` / `LowRankGaussianLoss`	`MixtureDensityLoss`, `NormalizingFlowLoss`	Prefer low-rank/full covariance when Gaussian is enough.
Noisy features / measurement error	`InputNoiseAugmentationLoss + GaussianCRPSLoss` / `MixtureDensityLoss` / `InputNoiseBinnedPDFLoss`	`FunctionalEIVLoss`, `StructuralEIVLoss`, `OrthogonalDistanceRegressionLoss`, `ForwardInversionRegressor`, `SplitMeasurementLoss`, `CorrectedScoreEIVLoss`	Start with explicit input-noise marginalization and test-time predictive averaging, then escalate to Jacobian-based EIV losses only if they clearly help. When attenuation bias dominates and labels are much more precise than the features, prefer the generative y->x inversion or a corrected-score loss; with repeat exposures use the split-measurement loss.
Noisy labels / label corruption	`WeightedHuberLoss`	`BaseEnsembleModel`, `ConformalLoss`	Prefer robust baselines before heavier methods.
Imbalanced / rare-target regression	`GaussianCRPSLoss` / `QuantileLoss + tail-slice evaluation`	`DensityConformal`, `FeatureDistributionSmoother`, `PrimeLoss`, `ConR`, `ContinuousGroupDRO`, `VIRRegressor`, `BalancedMSEBMC`, `RankSimLoss`	VIR adds evidential predictive uncertainty to imbalanced regression; use it when tail accuracy and calibrated uncertainty are both required. Astronomical benchmarks do not justify density weighting as default. Advanced research methods (DensityWeightedLoss, LDSLoss) should only be tried if coverage/calibration allow for tail gains. Representation-level methods (FDS, PRIME, ConR) and group-robust training (Group DRO over target bins) address feature collapse and worst-regime bias that output reweighting cannot fix.
Likelihood-ratio / simulation-based parameter estimation	`NeuralRatioLoss`	`ContrastiveFlowLoss`, `NormalizingFlowLoss`	Use NRE when a simulator can generate (parameters, observations) and the target is an MLE/MAP estimate, likelihood-ratio test or hypothesis ranking. Draw negative contexts from the simulation/prior distribution, and validate posterior weights and intervals with SBC/TARP before trusting them. ContrastiveFlowLoss is the flow-log-density variant; an NPE-style p(theta
Selection bias / covariate-dependent missing labels	`PropensityWeightedLoss`	`DensityWeightedLoss`	Estimate p(observed
Output constraints / monotonicity	`BoundedHead` / `NonNegativeHead` / `NonCrossingSort`	`SimplexHead`, `SpectralNormWrapper`	Apply structural constraints in the head before post-hoc calibration.
Post-hoc calibration transforms	`VarianceTemperatureScaler`	`IsotonicMeanCalibrator`, `PITCalibrator`, `VarianceRestoringCalibrator`	Fit transforms on a held-out calibration split. VarianceRestoringCalibrator matches the marginal distribution (tails/population variance) and is not a substitute for conditional mean/variance calibration.
Calibrated intervals with coverage guarantees	`ConformalLoss on top of a strong probabilistic backbone`	`QuantileLoss`, `MonteCarloConformal`, `DensityConformal`	Conformal gives coverage, not density estimation; keep CRPS/NLL reporting from the underlying predictive model.
Density-aware conformal under long-tail targets	`DensityConformal`	`PrevalenceAdjustedCP`, `MonteCarloConformal`	Prefer density/prevalence variants when tail-region coverage is a key objective; point-accuracy gains are not guaranteed.
Uncertain ground-truth / weak labels	`NoisyTargetGaussianNLL`	`OrdinalCrossEntropyLoss`, `PseudoLabelConsistencyLoss`, `PseudoLabelNLL`	Use NoisyTargetGaussianNLL for Gaussian label uncertainty and soft-bin / PMF supervision when target PDFs or intervals are the natural label form.
Semi-supervised regression	`PseudoLabelConsistencyLoss`	`PseudoLabelNLL`, `NoisyTargetGaussianNLL`	Use confidence-gated pseudo labels and keep a clean held-out evaluation split.
Target transforms for skewed / multiplicative-noise regression	`LogTransformLoss`	`BoxCoxTransformLoss`, `SqrtTransformLoss`, `YeoJohnsonTransformLoss`	Match transform support to target support before tuning model complexity.
Causal inference regression (ATE/CATE)	`dr_ate` / `dr_cate`	`PredictionPoweredInference`	Use cross-fitting and overlap diagnostics before interpreting treatment effects.
Population inference with few labels	`PredictionPoweredInference`	`ConformalLoss`, `QuantileLoss`	Use PPI for means/quantiles/regression coefficients with limited labels.
Ordinal / ordered targets	`CumulativeLinkLoss`	`OrdinalCrossEntropyLoss`	Prefer cumulative objectives when rank-distance errors matter.
Censored / interval-censored regression	`CensoredGaussianNLLLoss`	`AFTLoss`, `CensoredQuantileLoss`	Use censoring code 0/1/-1 and explicit interval bounds when available.
OOD scoring / selective prediction	`BaseEnsembleModel + OOD metrics`	`HeteroscedasticBatchEnsembleModel + OOD metrics`, `SWAG + OOD metrics`, `BayesianNeuralNetwork + OOD metrics`	Use multiple signals; no single OOD score is sufficient.

Method Capability Matrix (Peer Framing)¶

Legend:

Yes: directly supported
Partial: supported with conditions / composition
No: not a natural fit

Generated provenance: tools/render_method_catalog.py:render_family_capability_matrix_table Source artifacts: reports/method_catalog_latest.json Generated date: 2026-11-08

Method Family	Multi-target	Multimodal	Non-Gaussian	Epistemic	Aleatoric	Decomposition	Calibration	OOD Support	Imbalance	Noisy Features (EIV)
`bnn` (7)	yes	no	partial	yes	yes	yes	partial	partial	no	no
`calibration_transform` (7)	yes	no	partial	no	no	no	yes	partial	no	no
`causal` (2)	yes	no	partial	no	no	no	partial	partial	no	no
`censored` (3)	yes	no	yes	no	no	no	partial	partial	no	no
`conformal` (18)	yes	yes	yes	partial	no	no	yes	partial	yes	no
`constraints` (2)	yes	no	partial	no	no	no	partial	partial	no	no
`density_loss` (3)	yes	yes	yes	no	yes	no	partial	partial	no	no
`eiv` (16)	yes	yes	partial	no	no	no	yes	partial	no	yes
`ensemble` (9)	yes	yes	yes	yes	yes	yes	yes	yes	no	no
`evidential` (1)	yes	no	partial	yes	yes	yes	partial	partial	no	no
`expectile` (2)	yes	no	yes	no	no	no	partial	partial	no	no
`flow` (5)	yes	yes	yes	no	yes	partial	partial	partial	no	no
`flow_map` (1)	yes	yes	yes	no	yes	no	partial	partial	no	no
`flow_matching` (1)	yes	yes	yes	no	partial	no	partial	partial	no	no
`imbalanced_algorithm` (4)	yes	no	partial	no	no	no	partial	partial	yes	no
`imbalanced_loss` (11)	yes	no	partial	no	no	no	partial	partial	yes	no
`inference` (1)	yes	no	partial	no	no	no	partial	partial	no	no
`likelihood_ratio` (2)	partial	no	yes	partial	no	no	yes	partial	no	no
`mc_dropout` (1)	yes	no	partial	yes	partial	partial	partial	partial	no	no
`mdn` (1)	yes	yes	yes	no	yes	no	partial	partial	no	no
`multivariate` (1)	yes	no	partial	no	no	no	partial	partial	no	no
`observation_mapping` (3)	yes	no	partial	no	no	no	partial	partial	no	no
`ordinal` (2)	yes	no	yes	no	no	no	partial	partial	no	no
`point_loss` (1)	yes	no	no	no	no	no	partial	partial	no	no
`poisson` (4)	yes	no	yes	no	no	no	partial	partial	no	no
`poisson_gaussian` (4)	yes	yes	yes	no	no	no	partial	partial	no	no
`probabilistic_loss` (11)	yes	no	partial	no	yes	no	partial	partial	no	no
`quantile` (2)	yes	no	yes	no	no	no	yes	partial	no	no
`representation_regularizer` (1)	no	no	partial	no	no	no	no	partial	yes	no
`robust_algorithm` (1)	yes	no	partial	no	no	no	partial	partial	no	no
`robust_loss` (9)	yes	no	partial	no	no	no	partial	partial	no	no
`semi_supervised` (1)	yes	no	partial	no	no	no	partial	partial	no	no
`swag` (2)	yes	no	partial	yes	partial	partial	partial	partial	no	no
`tabular_network` (1)	yes	no	partial	no	partial	no	partial	partial	no	no
`target_transform` (5)	yes	no	partial	no	no	no	partial	partial	no	no
`test_time` (8)	yes	no	partial	partial	partial	partial	yes	partial	no	no
`tictac` (1)	yes	no	partial	no	yes	no	partial	partial	no	no
`tweedie` (4)	yes	no	yes	no	no	no	partial	partial	no	no
`uncertain_gt` (4)	yes	no	partial	no	yes	no	partial	partial	no	no

Code-Driven Catalog (Discovery API)¶

The tables above are task guidance. For scriptable filtering (and to reduce docs drift), use the public catalog metadata API:

import torchregress as tr

# Example: shortlist peer methods for OOD + epistemic signals
candidates = tr.method_catalog.list_methods(
    task_tag="ood",
    capability_filters={"epistemic": "yes"},
)

for row in candidates:
    print(row["name"], row["family"], row["maturity"])

Generated snapshot (from code): method_catalog_generated.md

Next step for operational tradeoffs: Choosing Methods by Constraint

Catalog-Backed Peer Method Snapshot (Generated)¶

Generated provenance: tools/render_method_catalog.py:render_method_matrix_generated_section Source artifacts: docs/reports/method_catalog_generated.md, reports/method_catalog_latest.json Generated date: 2026-11-08

Method	Family	Maturity	Multi-target	Multimodal	Epistemic	Aleatoric	Decomposition	Calibration	OOD
`BayesianNeuralNetwork`	`bnn`	`Available`	yes	no	yes	partial	partial	partial	partial
`HeteroscedasticBNN`	`bnn`	`Available`	yes	no	yes	yes	yes	partial	partial
`HeteroscedasticLaplaceRegressor`	`bnn`	`Available`	yes	no	yes	yes	yes	partial	partial
`IVON`	`bnn`	`Available`	yes	no	yes	no	no	partial	partial
`NaturalHeteroscedasticHead`	`bnn`	`Available`	yes	no	no	yes	no	partial	partial
`NaturalReparamHead`	`bnn`	`Available`	yes	no	no	yes	no	partial	partial
`VIDSRegressor`	`bnn`	`Available`	yes	no	yes	yes	no	partial	partial
`BinnedLabelShiftEstimator`	`calibration_transform`	`Strong`	yes	no	no	no	no	yes	partial
`CalibratedRegressor`	`calibration_transform`	`Available`	no	no	no	no	no	yes	partial
`IsotonicMeanCalibrator`	`calibration_transform`	`Available`	yes	no	no	no	no	yes	partial
`PITCalibrator`	`calibration_transform`	`Available`	yes	no	no	no	no	yes	partial
`SemiConformalCalibrator`	`calibration_transform`	`Strong`	yes	no	no	no	no	yes	partial
`VarianceRestoringCalibrator`	`calibration_transform`	`Available`	yes	no	no	no	no	yes	partial
`VarianceTemperatureScaler`	`calibration_transform`	`Available`	yes	no	no	no	no	yes	partial
`dr_ate`	`causal`	`Available`	yes	no	no	no	no	partial	partial
`dr_cate`	`causal`	`Available`	yes	no	no	no	no	partial	partial
`AFTLoss`	`censored`	`Available`	yes	no	no	no	no	partial	partial
`CensoredGaussianNLLLoss`	`censored`	`Available`	yes	no	no	no	no	partial	partial
`CensoredQuantileLoss`	`censored`	`Available`	yes	no	no	no	no	partial	partial
`CQR`	`conformal`	`Strong`	yes	no	no	no	no	yes	partial
`CTI`	`conformal`	`Available`	yes	no	no	no	no	yes	partial
`CVPlus`	`conformal`	`Available`	yes	no	no	no	no	yes	partial
`ConformalLoss`	`conformal`	`Core`	yes	no	no	no	no	yes	partial
`ConformalRegressor`	`conformal`	`Available`	no	no	no	no	no	yes	partial
`ConformalizedDensity`	`conformal`	`Available`	yes	no	no	no	no	yes	partial
`ContrastiveFlowConformal`	`conformal`	`Available`	yes	no	no	no	no	yes	partial
`DensityConformal`	`conformal`	`Available`	yes	no	no	no	no	yes	partial
`EnsembleBatchCP`	`conformal`	`Available`	yes	no	no	no	no	yes	partial
`MonteCarloConformal`	`conformal`	`Available`	yes	no	partial	no	no	yes	partial
`PrevalenceAdjustedCP`	`conformal`	`Available`	yes	no	no	no	no	yes	partial
`SLSConformal`	`conformal`	`Available`	yes	yes	no	no	no	yes	partial
`SplitConformal`	`conformal`	`Core`	yes	no	no	no	no	yes	partial
`UACQR`	`conformal`	`Available`	yes	no	no	no	no	yes	partial
`WaldoConformal`	`conformal`	`Available`	yes	no	no	no	no	yes	partial
`WaldoCriticalValueRegressor`	`conformal`	`Available`	yes	no	no	no	no	yes	partial
`WaldoDensityConformal`	`conformal`	`Available`	yes	no	no	no	no	yes	partial
`WaldoPosteriorCalibrator`	`conformal`	`Available`	yes	no	no	no	no	yes	partial
`BoundedHead`	`constraints`	`Available`	yes	no	no	no	no	partial	partial
`NonCrossingSort`	`constraints`	`Available`	yes	no	no	no	no	partial	partial
`BinnedDensityLoss`	`density_loss`	`Available`	no	yes	no	yes	no	partial	partial
`JointBinnedDensityLoss`	`density_loss`	`Available`	yes	yes	no	yes	no	partial	partial
`NoiseAwareDensityLoss`	`density_loss`	`Available`	yes	partial	no	yes	no	partial	partial
`CorrectedScoreEIVLoss`	`eiv`	`Available`	yes	no	no	no	no	partial	partial
`ErrorAwareFeatureEncoder`	`eiv`	`Available`	yes	no	no	no	no	partial	partial
`ForwardInversionRegressor`	`eiv`	`Available`	yes	no	no	no	no	partial	partial
`FunctionalEIVLoss`	`eiv`	`Available`	yes	no	no	no	no	partial	partial
`InputNoiseAugmentationLoss`	`eiv`	`Strong`	yes	no	no	no	no	partial	partial
`InputNoiseBinnedPDFLoss`	`eiv`	`Available`	yes	yes	no	no	no	partial	partial
`InputNoiseMDNLoss`	`eiv`	`Available`	yes	yes	no	no	no	partial	partial
`LatentMarginalizationLoss`	`eiv`	`Strong`	yes	no	no	no	no	partial	partial
`LatentNN`	`eiv`	`Available`	yes	no	no	no	no	partial	partial
`NoiseAwareRegressor`	`eiv`	`Available`	yes	no	no	no	no	partial	partial
`NoisyInputPredictor`	`eiv`	`Strong`	yes	no	no	no	no	partial	partial
`OrthogonalDistanceRegressionLoss`	`eiv`	`Available`	yes	no	no	no	no	partial	partial
`RegressionCalibration`	`eiv`	`Strong`	yes	no	no	no	no	yes	partial
`SIMEX`	`eiv`	`Strong`	yes	no	no	no	no	partial	partial
`SplitMeasurementLoss`	`eiv`	`Available`	yes	no	no	no	no	partial	partial
`StructuralEIVLoss`	`eiv`	`Available`	yes	no	no	no	no	partial	partial
`BaseEnsembleModel`	`ensemble`	`Core`	yes	no	yes	partial	partial	partial	yes
`BatchEnsembleRegressor`	`ensemble`	`Available`	yes	no	yes	partial	partial	partial	partial
`BinnedPDFEnsembleModel`	`ensemble`	`Available`	yes	no	yes	partial	partial	partial	partial
`CumulativeLinkEnsembleModel`	`ensemble`	`Available`	yes	no	yes	partial	partial	partial	partial
`DensityEnsemble`	`ensemble`	`Available`	yes	yes	yes	yes	yes	partial	partial
`HeteroscedasticBatchEnsembleModel`	`ensemble`	`Strong`	yes	no	yes	yes	yes	partial	partial
`HeteroscedasticEnsembleModel`	`ensemble`	`Strong`	yes	no	yes	yes	yes	partial	yes
`MDNEnsembleModel`	`ensemble`	`Available`	yes	no	yes	yes	yes	partial	partial
`calibrated_deep_ensemble`	`ensemble`	`Available`	no	no	yes	yes	partial	yes	partial
`EvidentialRegressionLoss`	`evidential`	`Available`	yes	no	yes	yes	yes	partial	partial
`ExpectileLoss`	`expectile`	`Available`	yes	no	no	no	no	partial	partial
`MultiExpectileLoss`	`expectile`	`Available`	yes	no	no	no	no	partial	partial
`ContrastiveFlowLoss`	`flow`	`Available`	yes	yes	no	yes	partial	partial	partial
`GaussianCopulaSplineHead`	`flow`	`Available`	yes	no	no	yes	no	partial	partial
`MarginalContrastiveDensity`	`flow`	`Available`	yes	no	no	yes	no	partial	partial
`NormalizingFlowLoss`	`flow`	`Available`	yes	yes	no	yes	partial	partial	partial
`SLSLoss`	`flow`	`Available`	yes	yes	no	no	no	partial	partial
`NormalizingFlowMap`	`flow_map`	`Available`	yes	yes	no	yes	no	partial	partial
`FlowMatchingMLP`	`flow_matching`	`Available`	yes	yes	no	partial	no	partial	partial
`ContinuousGroupDRO`	`imbalanced_algorithm`	`Available`	yes	no	no	no	no	partial	partial
`GroupDROTrainer`	`imbalanced_algorithm`	`Available`	yes	no	no	no	no	partial	partial
`VIRLoss`	`imbalanced_algorithm`	`Available`	yes	no	no	no	no	partial	partial
`VIRRegressor`	`imbalanced_algorithm`	`Available`	yes	no	no	no	no	partial	partial
`BalancedMSEBMC`	`imbalanced_loss`	`Available`	no	no	no	no	no	partial	partial
`BalancedMSELoss`	`imbalanced_loss`	`Available`	yes	no	no	no	no	partial	partial
`BinReweightedMSELoss`	`imbalanced_loss`	`Available`	yes	no	no	no	no	partial	partial
`ConR`	`imbalanced_loss`	`Available`	yes	no	no	no	no	partial	partial
`DensityWeightedLoss`	`imbalanced_loss`	`Available`	yes	no	no	no	no	partial	partial
`FeatureDistributionSmoother`	`imbalanced_loss`	`Available`	yes	no	no	no	no	partial	partial
`FocalRLoss`	`imbalanced_loss`	`Available`	yes	no	no	no	no	partial	partial
`LDSLoss`	`imbalanced_loss`	`Available`	yes	no	no	no	no	partial	partial
`PrimeLoss`	`imbalanced_loss`	`Available`	yes	no	no	no	no	partial	partial
`PropensityWeightedLoss`	`imbalanced_loss`	`Available`	yes	no	no	no	no	partial	partial
`RankSimLoss`	`imbalanced_loss`	`Available`	no	no	no	no	no	partial	partial
`PredictionPoweredInference`	`inference`	`Available`	yes	no	no	no	no	partial	partial
`NeuralRatioLoss`	`likelihood_ratio`	`Available`	partial	no	no	no	no	partial	partial
`WifiRatioEnsemble`	`likelihood_ratio`	`Available`	partial	no	partial	no	no	yes	partial
`MCDropoutWrapper`	`mc_dropout`	`Strong`	yes	no	yes	partial	partial	partial	partial
`MixtureDensityLoss`	`mdn`	`Available`	yes	yes	no	yes	no	partial	partial
`TaskAgnosticCorrelations`	`multivariate`	`Available`	yes	no	no	no	no	partial	partial
`CommonScaleSplineHead`	`observation_mapping`	`Available`	yes	no	no	no	no	partial	partial
`ConditionBiasHead`	`observation_mapping`	`Available`	yes	no	no	no	no	partial	partial
`RelativeSurveySplineHead`	`observation_mapping`	`Available`	yes	no	no	no	no	partial	partial
`CumulativeLinkLoss`	`ordinal`	`Available`	yes	no	no	no	no	partial	partial
`OrdinalCrossEntropyLoss`	`ordinal`	`Available`	yes	no	no	no	no	partial	partial
`WeightedMSELoss`	`point_loss`	`Core`	yes	no	no	no	no	partial	partial
`NegativeBinomialNLLLoss`	`poisson`	`Available`	yes	no	no	no	no	partial	partial
`PoissonDevianceLoss`	`poisson`	`Available`	yes	no	no	no	no	partial	partial
`PoissonLikelihoodRatioLoss`	`poisson`	`Available`	yes	no	no	no	no	partial	partial
`ZeroInflatedPoissonNLLLoss`	`poisson`	`Available`	yes	no	no	no	no	partial	partial
`EnhancedPoissonGaussianContaminationMixtureLoss`	`poisson_gaussian`	`Available`	yes	yes	no	no	no	partial	partial
`PoissonGaussianContaminationMixtureLoss`	`poisson_gaussian`	`Available`	yes	yes	no	no	no	partial	partial
`PoissonGaussianLikelihoodRatioLoss`	`poisson_gaussian`	`Available`	yes	no	no	no	no	partial	partial
`PoissonGaussianLoss`	`poisson_gaussian`	`Available`	yes	no	no	no	no	partial	partial
`BetaNLLLoss`	`probabilistic_loss`	`Available`	yes	no	no	yes	no	partial	partial
`FaithfulGaussianLoss`	`probabilistic_loss`	`Available`	yes	no	no	yes	no	partial	partial
`GaussianCRPSLoss`	`probabilistic_loss`	`Strong`	yes	no	no	yes	no	partial	partial
`GaussianLatentForwardObservationNLLLoss`	`probabilistic_loss`	`Available`	no	no	no	yes	no	partial	partial
`GaussianNLLLoss`	`probabilistic_loss`	`Core`	yes	no	no	yes	no	partial	partial
`GaussianWassersteinBoundLoss`	`probabilistic_loss`	`Available`	yes	no	no	partial	no	partial	partial
`LowRankGaussianLoss`	`probabilistic_loss`	`Strong`	yes	no	no	yes	no	partial	partial
`MarginalizedMultiSourceGaussianNLLLoss`	`probabilistic_loss`	`Available`	yes	no	no	yes	no	partial	partial
`MultivariateGaussianLoss`	`probabilistic_loss`	`Strong`	yes	no	no	yes	no	partial	partial
`NeighborhoodCovariancePseudoLabeler`	`probabilistic_loss`	`Available`	yes	no	no	partial	no	partial	partial
`RobustMultiSourceGaussianNLLLoss`	`probabilistic_loss`	`Available`	yes	no	no	yes	no	partial	partial
`MultiQuantileLoss`	`quantile`	`Strong`	yes	no	no	no	no	yes	partial
`QuantileLoss`	`quantile`	`Core`	yes	no	no	no	no	yes	partial
`UncertaintyAwareConR`	`representation_regularizer`	`Available`	no	no	no	no	no	no	partial
`iteratively_reweighted_fit`	`robust_algorithm`	`Strong`	yes	no	no	no	no	partial	partial
`AdaptiveRobustLoss`	`robust_loss`	`Available`	yes	no	no	no	no	partial	partial
`BarronLoss`	`robust_loss`	`Available`	yes	no	no	no	no	partial	partial
`CVaRLoss`	`robust_loss`	`Available`	yes	no	no	no	no	partial	partial
`CauchyLoss`	`robust_loss`	`Available`	yes	no	no	no	no	partial	partial
`CharbonnierLoss`	`robust_loss`	`Strong`	yes	no	no	no	no	partial	partial
`LogCoshLoss`	`robust_loss`	`Strong`	yes	no	no	no	no	partial	partial
`PseudoHuberLoss`	`robust_loss`	`Strong`	yes	no	no	no	no	partial	partial
`TukeyBiweightLoss`	`robust_loss`	`Available`	yes	no	no	no	no	partial	partial
`WeightedHuberLoss`	`robust_loss`	`Core`	yes	no	no	no	no	partial	partial
`TeacherStudentTrainer`	`semi_supervised`	`Strong`	yes	no	no	no	no	partial	partial
`MultiSWAG`	`swag`	`Available`	yes	no	yes	partial	partial	partial	partial
`SWAG`	`swag`	`Available`	yes	no	yes	partial	partial	partial	partial
`TabularMLP`	`tabular_network`	`Available`	yes	no	no	partial	no	partial	partial
`BoxCoxTransformLoss`	`target_transform`	`Available`	yes	no	no	no	no	partial	partial
`LogTransformLoss`	`target_transform`	`Available`	yes	no	no	no	no	partial	partial
`SqrtTransformLoss`	`target_transform`	`Available`	yes	no	no	no	no	partial	partial
`TransformedTargetLoss`	`target_transform`	`Available`	yes	no	no	no	no	partial	partial
`YeoJohnsonTransformLoss`	`target_transform`	`Available`	yes	no	no	no	no	partial	partial
`BayesianLinearHead`	`test_time`	`Available`	yes	no	partial	partial	partial	partial	no
`DelayedLabelResidualAdapter`	`test_time`	`Available`	yes	no	no	no	no	partial	partial
`OptimalTransportCoverageGap`	`test_time`	`Available`	yes	no	no	no	no	partial	partial
`RecursiveBayesianHead`	`test_time`	`Available`	yes	no	partial	partial	partial	partial	no
`ScoreCDFReweighter`	`test_time`	`Available`	yes	no	no	no	no	partial	partial
`WeightedConformalRegressionAdapter`	`test_time`	`Available`	yes	no	no	no	no	yes	partial
`WeightedSplitConformalAdapter`	`test_time`	`Available`	yes	no	no	no	no	partial	partial
`weighted_split_classification_predictive_batch`	`test_time`	`Available`	yes	no	no	no	no	partial	partial
`TaylorInducedCovarianceHead`	`tictac`	`Available`	yes	no	no	yes	no	partial	partial
`CompoundPoissonLoss`	`tweedie`	`Available`	yes	no	no	no	no	partial	partial
`GammaLoss`	`tweedie`	`Available`	yes	no	no	no	no	partial	partial
`InverseGaussianLoss`	`tweedie`	`Available`	yes	no	no	no	no	partial	partial
`TweedieLoss`	`tweedie`	`Available`	yes	no	no	no	no	partial	partial
`ConsistencyRegLoss`	`uncertain_gt`	`Available`	yes	no	no	no	no	partial	partial
`NoisyTargetGaussianNLL`	`uncertain_gt`	`Available`	yes	no	no	yes	no	partial	partial
`PseudoLabelConsistencyLoss`	`uncertain_gt`	`Available`	yes	no	no	no	no	partial	partial
`PseudoLabelNLL`	`uncertain_gt`	`Available`	yes	no	no	no	no	partial	partial

Peer-method check: SWAG, BayesianNeuralNetwork, MixtureDensityLoss

Generated Family Capability Matrix¶

Family	# Methods	Multi-target	Multimodal	Non-Gaussian	Epistemic	Aleatoric	Decomposition	Calibration	OOD	Imbalance	EIV
`bnn`	7	yes	no	partial	yes	yes	yes	partial	partial	no	no
`calibration_transform`	7	yes	no	partial	no	no	no	yes	partial	no	no
`causal`	2	yes	no	partial	no	no	no	partial	partial	no	no
`censored`	3	yes	no	yes	no	no	no	partial	partial	no	no
`conformal`	18	yes	yes	yes	partial	no	no	yes	partial	yes	no
`constraints`	2	yes	no	partial	no	no	no	partial	partial	no	no
`density_loss`	3	yes	yes	yes	no	yes	no	partial	partial	no	no
`eiv`	16	yes	yes	partial	no	no	no	yes	partial	no	yes
`ensemble`	9	yes	yes	yes	yes	yes	yes	yes	yes	no	no
`evidential`	1	yes	no	partial	yes	yes	yes	partial	partial	no	no
`expectile`	2	yes	no	yes	no	no	no	partial	partial	no	no
`flow`	5	yes	yes	yes	no	yes	partial	partial	partial	no	no
`flow_map`	1	yes	yes	yes	no	yes	no	partial	partial	no	no
`flow_matching`	1	yes	yes	yes	no	partial	no	partial	partial	no	no
`imbalanced_algorithm`	4	yes	no	partial	no	no	no	partial	partial	yes	no
`imbalanced_loss`	11	yes	no	partial	no	no	no	partial	partial	yes	no
`inference`	1	yes	no	partial	no	no	no	partial	partial	no	no
`likelihood_ratio`	2	partial	no	yes	partial	no	no	yes	partial	no	no
`mc_dropout`	1	yes	no	partial	yes	partial	partial	partial	partial	no	no
`mdn`	1	yes	yes	yes	no	yes	no	partial	partial	no	no
`multivariate`	1	yes	no	partial	no	no	no	partial	partial	no	no
`observation_mapping`	3	yes	no	partial	no	no	no	partial	partial	no	no
`ordinal`	2	yes	no	yes	no	no	no	partial	partial	no	no
`point_loss`	1	yes	no	no	no	no	no	partial	partial	no	no
`poisson`	4	yes	no	yes	no	no	no	partial	partial	no	no
`poisson_gaussian`	4	yes	yes	yes	no	no	no	partial	partial	no	no
`probabilistic_loss`	11	yes	no	partial	no	yes	no	partial	partial	no	no
`quantile`	2	yes	no	yes	no	no	no	yes	partial	no	no
`representation_regularizer`	1	no	no	partial	no	no	no	no	partial	yes	no
`robust_algorithm`	1	yes	no	partial	no	no	no	partial	partial	no	no
`robust_loss`	9	yes	no	partial	no	no	no	partial	partial	no	no
`semi_supervised`	1	yes	no	partial	no	no	no	partial	partial	no	no
`swag`	2	yes	no	partial	yes	partial	partial	partial	partial	no	no
`tabular_network`	1	yes	no	partial	no	partial	no	partial	partial	no	no
`target_transform`	5	yes	no	partial	no	no	no	partial	partial	no	no
`test_time`	8	yes	no	partial	partial	partial	partial	yes	partial	no	no
`tictac`	1	yes	no	partial	no	yes	no	partial	partial	no	no
`tweedie`	4	yes	no	yes	no	no	no	partial	partial	no	no
`uncertain_gt`	4	yes	no	partial	no	yes	no	partial	partial	no	no

Generated Hard-Task Shortlists¶

Need	Catalog Filter (conceptual)	Suggested Methods
OOD + epistemic signals	`task_tag='ood'` + `epistemic=yes`	`BayesianNeuralNetwork`, `HeteroscedasticBNN`, `BaseEnsembleModel`, `BatchEnsembleRegressor`, `HeteroscedasticBatchEnsembleModel`, `HeteroscedasticEnsembleModel`, `MultiSWAG`, `SWAG`
Coverage / calibration	`calibration=yes`	`BinnedLabelShiftEstimator`, `CalibratedRegressor`, `IsotonicMeanCalibrator`, `PITCalibrator`, `SemiConformalCalibrator`, `VarianceRestoringCalibrator`, `VarianceTemperatureScaler`, `CQR`, `CTI`, `CVPlus`, `ConformalLoss`, `ConformalRegressor`, `ConformalizedDensity`, `ContrastiveFlowConformal`, `DensityConformal`, `EnsembleBatchCP`, `MonteCarloConformal`, `PrevalenceAdjustedCP`, `SLSConformal`, `SplitConformal`, `UACQR`, `WaldoConformal`, `WaldoCriticalValueRegressor`, `WaldoDensityConformal`, `WaldoPosteriorCalibrator`, `RegressionCalibration`, `calibrated_deep_ensemble`, `WifiRatioEnsemble`, `MultiQuantileLoss`, `QuantileLoss`, `WeightedConformalRegressionAdapter`
Multimodal targets	`multimodal=yes`	`SLSConformal`, `BinnedDensityLoss`, `JointBinnedDensityLoss`, `InputNoiseBinnedPDFLoss`, `InputNoiseMDNLoss`, `DensityEnsemble`, `ContrastiveFlowLoss`, `NormalizingFlowLoss`, `SLSLoss`, `NormalizingFlowMap`, `FlowMatchingMLP`, `MixtureDensityLoss`, `EnhancedPoissonGaussianContaminationMixtureLoss`, `PoissonGaussianContaminationMixtureLoss`
Imbalanced / rare targets	`imbalance=yes`	`DensityConformal`, `PrevalenceAdjustedCP`, `ContinuousGroupDRO`, `GroupDROTrainer`, `VIRLoss`, `VIRRegressor`, `BalancedMSEBMC`, `BalancedMSELoss`, `BinReweightedMSELoss`, `ConR`, `DensityWeightedLoss`, `FeatureDistributionSmoother`, `FocalRLoss`, `LDSLoss`, `PrimeLoss`, `PropensityWeightedLoss`, `RankSimLoss`, `UncertaintyAwareConR`
Noisy features / EIV	`noisy_features_eiv=yes`	`CorrectedScoreEIVLoss`, `ErrorAwareFeatureEncoder`, `ForwardInversionRegressor`, `FunctionalEIVLoss`, `InputNoiseAugmentationLoss`, `InputNoiseBinnedPDFLoss`, `InputNoiseMDNLoss`, `LatentMarginalizationLoss`, `LatentNN`, `NoiseAwareRegressor`, `NoisyInputPredictor`, `OrthogonalDistanceRegressionLoss`, `RegressionCalibration`, `SIMEX`, `SplitMeasurementLoss`, `StructuralEIVLoss`

Evidence-Based Maturity Labels (Current Audit Guidance)¶

These labels are intentionally evidence-based and implementation-specific.

Label	Meaning
`Core`	Stable API pattern, direct tests, docs coverage, examples
`Strong`	Good coverage and examples, some edge-case/test debt remains
`Available`	Implemented and documented, but fewer examples/benchmarks or less direct test depth
`Advanced`	Powerful but needs stronger decision-grade examples/benchmarks for adoption confidence

Current high-level guidance (family-level, not absolute):

Core: robust point losses, Gaussian losses, basic metrics, conformal basics, deep ensembles
Strong: calibration/OOD metrics, imbalanced losses, multi-target Gaussian variants
Available: SWAG, BNN, MDN, normalizing-flow losses, EIV losses
Advanced: flow-based multi-target comparisons, combined decomposition + calibration benchmarks

Evidence and Tradeoffs (What To Check Before Committing)¶

Use this matrix to shortlist methods, then validate the shortlist against examples and benchmarks:

Comparison examples: examples/comprehensive_comparison.py, examples/loss_comparison.py --suite comprehensive
Hard-task examples: examples/imbalanced_regression.py, examples/conformal_regression_example.py --mode sweep, examples/normalizing_flows_multitarget.py
Comparative coverage tracker: Comparative Evidence Matrix
Performance smoke/sweep baselines: tools/benchmark_smoke.py and reports under reports/benchmark_thresholds/

Key tradeoffs and common failure modes:

Conformal: strong coverage guarantees, but interval width can become large under shift or poor base models; does not provide epistemic/aleatoric decomposition.
MDN: strong multimodal baseline, but can be sensitive to component count and optimization; compare interval/calibration behavior, not just NLL.
Normalizing flows: flexible and expressive, but higher implementation/runtime complexity and optional dependency requirements; use when Gaussian/MDN miss important structure.
SWAG / BNN: useful epistemic signals, but calibration and runtime quality depend heavily on training protocol and posterior approximation choices; validate on your OOD and calibration metrics, not only in-distribution error.
Deep / heteroscedastic ensembles: strong defaults for uncertainty, but training and inference costs scale with ensemble size; benchmark with your latency budget.

Decision Workflow (Fast Path)¶

Generated provenance: tools/render_method_catalog.py:render_decision_workflow Source artifacts: reports/method_catalog_latest.json Generated date: 2026-11-08

Need coverage guarantees? Use ConformalLoss (split / CQR / ACI patterns). Caveat: Coverage guarantees are not uncertainty decomposition.
Need epistemic + aleatoric decomposition? Use HeteroscedasticEnsembleModel. Alternatives: HeteroscedasticBNN, MDNEnsembleModel, HeteroscedasticBatchEnsembleModel. Caveat: Requires variance/distribution modeling and stronger compute budget.
Need multimodal outputs? Use MixtureDensityLoss. Alternatives: NormalizingFlowLoss. Caveat: Move to flows when MDN component count/training stability is the bottleneck.
Have noisy features / measurement error? Use InputNoiseAugmentationLoss + GaussianCRPSLoss / MixtureDensityLoss / InputNoiseBinnedPDFLoss. Alternatives: FunctionalEIVLoss / StructuralEIVLoss / OrthogonalDistanceRegressionLoss. Caveat: Use the simpler explicit input-noise path first, including test-time predictive averaging; Jacobian-style EIV losses are more fragile and need careful benchmarking.
Have imbalanced tails / rare targets? Use GaussianCRPSLoss / QuantileLoss + tail-slice evaluation. Alternatives: DensityConformal, DensityWeightedLoss, LDSLoss. Caveat: Density-aware weighting is not yet a universally strong default on astronomical benchmarks.
Need OOD scoring / selective prediction under a latency budget? Use BaseEnsembleModel + OOD metrics. Alternatives: HeteroscedasticBatchEnsembleModel + OOD metrics, SWAG + OOD metrics, BayesianNeuralNetwork + OOD metrics, MCDropoutWrapper. Caveat: Use multiple signals and benchmark runtime against deployment latency targets.
Are labels uncertain or weak (noisy targets, pseudo-labels, partial trust)? Use NoisyTargetGaussianNLL. Alternatives: OrdinalCrossEntropyLoss, PseudoLabelNLL, ConsistencyRegLoss. Caveat: Retain held-out clean-label evaluation where available to avoid self-confirming loops.
Need treatment-effect estimation under confounding (ATE/CATE)? Use dr_ate / dr_cate. Alternatives: PredictionPoweredInference. Caveat: Check overlap/ESS diagnostics and avoid causal claims under severe positivity violations.

Important Clarification: Conformal vs Uncertainty Decomposition¶

Conformal prediction provides interval coverage guarantees.

It does not separate epistemic and aleatoric uncertainty.

Use conformal for calibrated intervals. Use heteroscedastic ensemble-style approaches when you need explicit decomposition. Use MixtureDensityLoss, NormalizingFlowLoss, QuantileLoss, or binned-PDF heads when you need richer predictive distributions. See Uncertainty Decomposition for the difference, including quantile ensembles.