Technical Summary: Decision-Aligned Evaluation of Uncertainty Quantification

Problem Statement

Uncertainty quantification (UQ) in machine learning is critical for reliable decision-making in safety-critical domains. However, the evaluation of UQ methods currently relies on generic surrogate metrics such as the negative log-likelihood (NLL), Brier score (BS), and expected calibration error (ECE). A fundamental disconnect exists: high performance on these standard metrics does not necessarily translate to high utility in downstream decision problems. The authors argue that current evaluation protocols risk optimizing for "wrong" downstream worlds because these metrics implicitly encode specific, often pathological, prior beliefs about decision costs or fail to align with decision utilities entirely.

Methodology

The paper introduces a formal framework called decision-alignment to analyze the relationship between evaluation metrics and downstream decision utilities.

1. Decision-Alignment Criterion

A metric $M$ is defined as decision-aligned with respect to a family of decision utilities $\{U_\theta\}_{\theta \in \Theta}$ if there exists a prior distribution $\pi$ over the utility parameters $\theta$ such that the metric value is a strictly increasing transformation of the expected utility under that prior. Formally:
$$h_y(M_y(f)) = \int_{\Theta} -U_{\theta, y}(f) \pi(\theta) d\theta$$
where $h_y$ is a strictly increasing function. This implies that a decision-aligned metric ranks models exactly as the expected utility would under the implicit prior $\pi$.

2. Analysis of Common Metrics

The authors apply this framework to standard UQ metrics across classification and regression tasks, considering decision families such as:

• Binary Decisions: Cost-sensitive classification (e.g., medical diagnosis).
• Top-k Selection: Selecting a subset of high-confidence predictions (e.g., recommender systems).
• Selective Prediction: Deciding whether to predict or abstain based on uncertainty (e.g., deferring to a human expert).

The analysis reveals that:

• Coordinate-Dependent Metrics: Metrics like ECE, MCE, R-AUC, and E-Det are not decision-aligned for pointwise-separable utility families because they couple predictions across the dataset, violating the separability required for alignment with standard decision utilities.
• Proper Scoring Rules (PSRs): While metrics like NLL and BS are proper scoring rules, they are only decision-aligned under specific, often pathological priors. For example:
  • NLL implies a prior $\pi(c) \propto 1/(c(1-c))$, placing unbounded mass on extreme cost ratios where one error type is essentially free.
  • Accuracy (Acc) implies a degenerate prior assuming perfectly symmetric costs.
  • BS implies a uniform prior, which may not reflect real-world cost asymmetries.
• Pathological Priors: The implicit priors of standard metrics often place mass on regions where uncertainty quantification is least needed (e.g., trivial decision boundaries) or assume unrealistic cost structures.

3. Proposed Solution: Prior-Weighted Utility (PWU) Metrics

To address these misalignments, the authors propose Prior-Weighted Utility (PWU) metrics. These are a class of proper scoring rules constructed by explicitly integrating a downstream utility function against a plausible prior $\pi$ that reflects realistic decision contexts:
$$M_\pi(f, y) = \int_{\Theta} -U_\theta(f, y) \pi(\theta) d\theta$$
Unlike standard metrics, PWU metrics allow researchers to explicitly define the decision context (via the utility family) and the belief about decision parameters (via the prior), ensuring the metric is inherently decision-aligned.

Key Contributions

1. Decision-Alignment Framework: A formal criterion to determine whether a UQ metric aligns with downstream utilities, revealing that common metrics either encode pathological priors or no coherent decision belief.
2. Theoretical Diagnosis: A systematic analysis showing that coordinate-dependent metrics (ECE, R-AUC, etc.) are fundamentally misaligned with pointwise-separable decision problems, and that even coordinate-independent PSRs (NLL, BS) often rely on implausible implicit priors.
3. Prior-Weighted Utility Metrics: The introduction of a new class of metrics that are by definition decision-aligned. These metrics provide a principled way to evaluate UQ methods based on their value for specific downstream decision scenarios.
4. Empirical Validation: Extensive experiments on benchmark datasets and real-world case studies demonstrating that PWU metrics consistently align with realized decision utility, whereas conventional metrics often fail to do so.

Results

The authors conducted experiments on 10 binary classification and 10 regression models across five datasets, evaluating alignment using Kendall's $\tau$ between metric rankings and utility rankings.

• Benchmark Datasets: In controlled experiments, PWU metrics consistently achieved the highest alignment with their corresponding utility families. In contrast, conventional metrics (NLL, BS, ECE, etc.) showed weak or negative alignment, particularly in Top-k selection and selective prediction tasks.
• Real-World Case Studies:
  • Electricity Market Bidding: In a wind farm bidding scenario, conventional metrics showed no stable positive alignment with bidding utility, while PWU metrics (specifically those for selective prediction) exhibited strong alignment.
  • Credit Approval & P2P Lending: In financial decision tasks with customer-dependent costs, PWU metrics robustly aligned with decision utility, even when the specific decision parameters in the case study differed from the priors used to construct the PWU metrics.
• Robustness: Sensitivity analyses demonstrated that PWU metrics remain robust to prior misspecification. Even when the prior was significantly perturbed, PWU metrics maintained strong positive utility alignment, outperforming conventional metrics which rely on fixed, pathological implicit priors.

Significance and Claims

The paper claims that the dominant practice of evaluating UQ with generic metrics can be systematically misleading when the goal is to support downstream decisions. The authors argue that:

1. Current Benchmarks are Flawed: Standard metrics silently optimize for pathological decision worlds, potentially promoting models that perform well on NLL or ECE but fail in practical, cost-sensitive applications.
2. Explicit Priors are Necessary: Meaningful progress in probabilistic ML requires moving away from implicit assumptions. By explicitly stating the prior $\pi$ and the utility family, researchers can ensure that UQ benchmarks evaluate methods that enable reliable decision-making.
3. A Principled Extension: PWU metrics offer a principled extension of existing evaluation protocols. They do not replace proper scoring rules but refine them to be decision-relevant.

The authors conclude that future UQ benchmarking should report a variety of PWU metrics alongside conventional ones to cover a range of plausible decision contexts, thereby surfacing methods that are truly useful for downstream applications.