PaperlayerView exact revision on arXiv

Source-linked AI summary

Cost-Aware Post-Hoc Deferral Under Calibration and Shift: An Environmental AI Case Study

Haoran Yu, Lifei Liu, Danping Zhang

arXiv:2609.09235v1cs.SEcs.LG

TL;DR

EcoTrust asks when post-hoc deferral around a frozen environmental classifier should rely on calibrated confidence, learned error risk, or a support-triggered fallback. It combines multi-signal risk estimation with cost-aware review decisions and finds that richer signals do not reliably reduce cost or transfer as case-level rankings in this controlled task.

  • Problem

    The paper examines whether richer uncertainty signals improve deployment cost over calibration-matched, class-aware confidence when errors, review quality, quotas, and extrapolation interact.

  • Method

    EcoTrust combines six-signal post-hoc error-risk estimation, calibration, class-asymmetric cost-aware deferral, reviewer accuracy, budgeted review, and an optional support gate.

  • Results

    Across the studied settings, learned risk does not reliably beat calibrated confidence: Chow costs 0.416 versus 0.567 per day on the primary backend, and learned risk wins on six of 13 alternatives.

  • Takeaways & Limitations

    For the well-calibrated backend studied, use Chow and consider learned risk only after validating lower cost on temporal calibration data.

  • Takeaways & Limitations

    The evidence is a controlled case study using 1,195 rows and ten stations within one task domain, and additional tasks and real reviewers remain necessary.

Abstract

from arXiv · show

Choosing a deferral policy for a frozen classifier requires more than ranking uncertain cases: confidence may be miscalibrated, errors have unequal costs, reviewers can err, and deployment data can leave calibration support. We study these interactions through EcoTrust, a post-hoc framework that compares automatic action with review using a six-group error-risk estimator, class-asymmetric costs, reviewer accuracy, and an optional support gate. On a Columbia River thermal-stress testbed, the learned estimator improves error-ranking area under the receiver operating characteristic curve from 0.869 to 0.889, but Chow's confidence rule has lower in-distribution cost (0.416 versus 0.567 per day). Across 12 off-the-shelf backends, learned risk and a calibration-matched, class-aware confidence estimator each beat raw Chow on six; a paired year-block bootstrap does not resolve their mean cost difference. In transfer to ten river stations, the gate flags every case and becomes an always-review fallback, attaining the lowest cost on eight stations only when review is perfect and unconstrained. These results characterize decision boundaries on one controlled task: richer risk signals do not reliably improve on calibrated confidence, and detected extrapolation does not imply transferable case-level ranking.

I. INTRODUCTION

EcoTrust evaluates post-hoc deferral for a frozen classifier when confidence, asymmetric costs, reviewer accuracy, workload limits, and distributional support jointly determine deployment decisions. It compares learned multi-signal risk with calibrated confidence and separates fallback behavior from transferable ranking.

  • I. INTRODUCTION: The Columbia River case assigns a missed stress day a cost of 100, a false alarm 3, and expert review 1, while limiting review workload.The same policy must also specify behavior outside calibration support.
  • I. INTRODUCTION: The unresolved question is whether richer uncertainty signals lower deployment cost than calibration-matched, class-aware confidence under imperfect review, quotas, and detected extrapolation.The study treats this as an empirical characterization around a frozen predictor rather than a new Bayes decision rule.
  • I. INTRODUCTION: The framework’s contributions are a cost-aware formulation linking asymmetric errors, reviewer quality, quotas, and extrapolation fallback, plus deployment comparisons across ranking and fallback behavior.These comparisons are designed to distinguish an always-review fallback from transferable case-level risk ranking.
  • I. INTRODUCTION: EcoTrust compares automatic action with review using six uncertainty signals, asymmetric error costs, reviewer accuracy, and an optional support fallback.The framework also supports budgeted review by ranking cases according to expected saving.

II. RELATED WORK

Prior work covers rejection, selective classification, learning to defer, post-hoc error estimation, and confidence sufficiency. EcoTrust narrows the comparison to frozen-backend deferral settings with class-dependent errors and detected extrapolation.

  • II. RELATED WORK: Earlier methods route uncertain predictions away from automatic action, jointly learn rejectors with classifiers or experts, or estimate fixed-model error post hoc.The related literature includes Chow’s rule, selective classification, learning to defer, and post-hoc confidence and density-ratio approaches.
  • II. RELATED WORK: The deployment-framework table summarizes the selected deferral methods and their settings.
  • II. RELATED WORK: An explicit fallback is defined as a pre-specified action after detected extrapolation, not as general robustness to distribution shift.
  • II. RELATED WORK: EcoTrust compares raw confidence, class-aware calibrated confidence, and multi-signal error risk for one reviewer under asymmetric consequences and detected extrapolation.Its scope is a narrower deployment comparison rather than a replacement for the neighboring methods.

B. Uncertainty Quantification and Environmental AI

The study holds an environmental predictor fixed and evaluates the downstream governance decision for thermal-stress predictions. Its policy minimizes expected deployment cost while accounting for final-decision errors, reviewer accuracy, and review budgets.

  • B. Uncertainty Quantification and Environmental AI: The environmental input uses NOAA weather and USGS hydrological features to predict a binary Chinook salmon thermal-stress threshold.Water temperature, month, and day-of-year are excluded from the model features.
  • B. Uncertainty Quantification and Environmental AI: A frozen predictor emits a probability and binary decision, while a governance policy chooses automatic action or review per case.The predictor remains fixed throughout governance training and evaluation.
  • B. Uncertainty Quantification and Environmental AI: Deployment quality is expected cost, where false negatives and false positives count in the final decision and reviewed cases can still incur consequence costs when the reviewer errs.The formulation also includes reviewer accuracy and an optional review budget.

IV. ECOTRUST DEPLOYMENT FRAMEWORK

EcoTrust estimates model error from six uncertainty groups, calibrates that estimate, and applies a cost-aware rule to choose automatic action or review. Budgeted review ranks cases by expected saving, while reviewer quality determines when escalation is beneficial.

  • A. Stage 1: The Meta Risk Estimator: EcoTrust’s meta-risk model maps confidence, entropy, ensemble disagreement, agent conflict, conformal quantities, and training-support distance to estimated prediction error.It is fitted by maximum likelihood on held-out data without recorded review decisions.
  • B. Stage 2: The Cost-Aware Decision Rule: The deployed decision compares estimated automatic error cost with review cost, assuming calibrated confidence and reviewer errors independent of the label.The rule chooses the lower-cost action and is applied with estimated risk and confidence.
  • B. Stage 2: The Cost-Aware Decision Rule: At error-cost ratio 100:3:1, the asymmetric thresholds are 0.01 for predicted-safe cases and 0.333 for predicted-stress cases.A single class-independent threshold cannot represent this asymmetry.
  • B. Stage 2: The Cost-Aware Decision Rule: Under a review budget, the optimal policy defers the cases with the largest positive expected savings, producing nested escalation sets and a convex cost–budget frontier.The framework uses an oracle diagnostic for the frontier and fixed-rate experiments for deployable rankings.
  • B. Stage 2: The Cost-Aware Decision Rule: Reviewer accuracy must exceed the model’s accuracy on the escalated set for deferral to improve expected accuracy.The measured breakeven lies between reviewer accuracies 0.90 and 0.95, around the predicted 0.9305.

C. Support Gate and Detected Extrapolation

EcoTrust uses a support gate to raise estimated risk outside the training distribution and trigger a predetermined review fallback. This safeguard is conditional rather than a general guarantee of robustness under distribution shift.

  • Support-gate design: The support gate raises deployed risk to maximum binary uncertainty for cases beyond the training-support threshold, encouraging review without target-domain labels.The gate uses a squared Mahalanobis distance and the (1−ϵ) training quantile, with ϵ=0.05.
  • Fallback guarantee: With a perfect reviewer, the fallback defers out-of-support cases when their error cost exceeds twice the review cost.Conditional on deferral, cost is bounded by ch + (1−a) max(cFN,cFP) and equals ch when review is perfect.
  • Deployment procedure: At deployment, EcoTrust fits cross-validated risk and support models, then routes a case to review when expected automatic error cost exceeds review cost.A budgeted variant instead defers the top-B fraction by expected saving.
  • Scope and limitation: The fallback removes corrupted model confidence from the invocation decision, but remains conditional and does not guarantee robustness to concept shift or undetected covariate shift.Its formal properties depend on the risk estimate and support detector behaving as intended under deployment conditions.

V. EXPERIMENTAL SETUP

The experiments use a temporally split Columbia River thermal-stress dataset and a fixed ensemble classifier, then compare governance policies under a stated asymmetric-cost scenario. EcoTrust is implemented as a cross-fitted, calibrated logistic meta-risk estimator alongside several baselines.

  • Data: The dataset combines NOAA weather and USGS hydrology observations from 1996–2024 for August–November migration-season evaluation at the Columbia River at The Dalles.Days without water temperature are dropped because their labels are undefined.
  • Splits: The test block contains 1,208 post-2014 cases, while training and calibration use earlier temporal blocks, preventing later dates from informing learned auxiliary signals.The frozen ensemble reaches 92.7% test accuracy.
  • Costs: The primary cost ratio is cFN:cFP:ch = 100:3:1, making missed stress far more expensive than false alarms while assigning one review unit per expert review.The ratio is illustrative, and cFN is swept from 1 to 500 in a sensitivity analysis.
  • Meta-risk estimator: EcoTrust uses an ℓ2-regularized logistic regression on standardized, median-imputed six-group signals with cross-fitted Platt scaling, while only the frozen predictor changes across backends.The estimator is fit without test-set hyperparameter tuning and is reused across the backend comparisons.
  • Baselines: The comparison includes automatic, random, heuristic, selective-classification, Chow, raw-feature, and post-hoc L2D baselines using the same frozen predictor.The L2D variants are frozen-backend surrogates rather than full joint classifier–rejector reproductions.

VI. MAIN COMPARISON

Chow is the strongest in-distribution policy despite EcoTrust’s slightly better error ranking. The comparison shows that ranking quality and deployment cost can diverge when review carries cost and policies review different fractions of cases.

  • Heuristics: 1.244 cost at 74.8% escalation for B1 and 0.814 cost at 55.1% review for B2 are both worse than the fully automatic cost-sensitive threshold’s 0.596.The heuristic policies therefore do not justify their tuning in this comparison.
  • Main policy comparison: 0.416 per day for Chow beats EcoTrust’s 0.567 in-distribution cost, while Chow reviews 31.0% of days versus EcoTrust’s 44.0%.The two post-hoc L2D variants cost 1.048 and 1.310, respectively.
  • Risk ranking: 0.889 AUC and 0.056 Brier for learned risk slightly improve on confidence’s 0.869 AUC and 0.058 Brier, yet better error ranking does not lower decision cost.The raw-feature rejector reaches 0.861 AUC and 0.061 Brier.
  • Coverage versus cost: 93.2% error coverage from B1 requires reviewing nearly three-quarters of cases, illustrating why error coverage alone is insufficient when review is charged.The figure contrasts coverage against escalation rate and deployment cost.

A. Which Signals Carry the Risk?

No individual signal is indispensable to EcoTrust’s risk model under the leave-one-out analysis. The full pre-specified signal bank is retained rather than selected using test cost.

  • Leave-one-out analysis: 0.011 maximum cost change and 0.015 maximum AUC change result from removing any one signal group.Across all 63 non-empty subsets, no removal increases cost, while entropy and conformal information contribute most to AUC.

VII. WHEN DOES A LEARNED RISK MODEL PAY?

The study tests whether a separately learned error-risk model adds deployment value beyond calibration-matched, class-aware confidence. Across backends, its benefits are associated with calibration but do not reliably exceed supervised recalibration or lower mean cost.

  • Learned risk wins on six of 13 real backends, while the ensemble costs 0.567 versus 0.416 for Chow per day.It reduces cost on the deliberately over-confident control from 1.478 to 0.566.
  • Calibration partly predicts learned-risk value, with Pearson correlation 0.278 and Spearman correlation 0.432 across the backend bank.The rank correlation’s year-block bootstrap interval includes zero, so miscalibration is treated as a diagnostic rather than a sufficient condition.
  • Removing any one signal changes test cost by at most 0.011 and risk AUC by at most 0.015, indicating no individually load-bearing signal.The leave-one-out result comes from the broader subset analysis summarized in Fig. 3.
  • The full six-signal bank is cheaper on seven of 12 backends than class-aware confidence, but its mean-cost interval crosses zero.The comparison uses paired year-block resampling and the primary cost setting with a perfect simulated reviewer.
  • Calibration-matched evaluation attributes the full bank’s advantage over raw confidence largely to supervised, class-aware recalibration rather than a reliably beneficial six-signal mechanism.The recalibrator uses the same Platt/logistic machinery, rows, folds, and support gate while restricting inputs to confidence-derived features.

VIII. COST SENSITIVITY AND REVIEW BUDGETS

Deployment cost changes with miss severity, review quotas, reviewer accuracy, and extrapolation. EcoTrust can outperform automation when misses are costly, but fixed-rate rankings vary, poor reviewers reduce benefits, and transfer gates may collapse to full review.

  • At cF N = 10, EcoTrust is 19.2% cheaper than automation, and at cF N = 100 it is 78.4% cheaper.When misses are cheap, review is not worthwhile: at cF N = 5, EcoTrust costs 0.275 versus 0.268 for automation; Chow remains cheaper at the primary point.
  • No ranking dominates at exact review rates: the raw-feature rejector is best at 5–10%, Chow at 15–20%, and selective classification at 30–50%.The hand-weighted composite is dominated at every displayed review rate, while forced exactly-k comparisons produce nonmonotone curves.
  • Reviewer-awareness lowers cost under noise but cannot compensate for a poor reviewer: at a = 0.70, cost falls from 1.994 to 1.563 while accuracy remains 0.842.The reviewer-aware policy also reduces escalation from 44.0% to 36.8%.
  • The transfer gate marks every day at every station out-of-support, becoming the always-review baseline at cost 1 with a perfect reviewer.This is cheapest on eight sites, but the result verifies a conservative fallback rather than transfer competence.
  • Under fixed review quotas, EcoTrust and Chow are nearly indistinguishable under shift, and the gate’s ranking does not recover transferable case-level risk.At a = 0.95 and B = 0.50, EcoTrust costs 1.773 versus 1.865 for automation and 1.778 for Chow; at B = 0.30, OvA is best at 1.694.
  • Mahalanobis gates flag all ten transfer stations, while increasing in-distribution flagging from 1.2% to 16.6% raises cost from 0.565 to 0.593.The sweep exposes a review-load trade-off rather than evidence of transferable ranking quality.

XI. DISCUSSION

EcoTrust characterizes deployment choices rather than establishing algorithmic dominance or general shift robustness. The study recommends Chow for the well-calibrated in-distribution backend, while treating learned risk and support gating as conditional choices requiring validation.

  • 94.1% versus 92.7% accuracy: a month-only logistic regression exceeds the ensemble, confirming that the study targets deferral decisions rather than raw prediction accuracy.
  • The study is limited by a 1,195-row meta-risk model, ten stations that broaden geography but not task domain, and simulated class-independent reviewer behavior.The gate detects covariate rather than concept shift and reviews every transfer case.
  • The support gate flags every transfer case and becomes always-review, attaining the lowest cost on eight stations only with a perfect reviewer and unconstrained review.L2D-OvA is cheaper on Clearwater and Snake.
  • Chow is cheaper for the primary in-distribution backend, while learned risk wins on six of 13 alternatives.A calibration-matched test does not resolve the mean cost difference between the full signal bank and a class-aware recalibrator.
Loading 2609.09235v1…