Utility-Directed Conformal Prediction: A Decision-Aware Framework for Actionable Uncertainty Quantification

TL;DR

This work extends conformal prediction by incorporating downstream utility into the construction of prediction sets, bridging model-agnostic uncertainty quantification with decision-focused objectives. It introduces two families of methods: separable-loss approaches with penalized non-conformity scores and non-separable-loss approaches via optimization-based set construction, both with formal coverage guarantees. The methods are validated on standard vision datasets and a dermatology case study, demonstrating substantially lower decision costs while preserving marginal coverage, and showing robustness to model quality and data scarcity. The framework enables hierarchical or domain-aligned interpretation of prediction sets, aiding high-stakes decision-making in fields like healthcare.

Abstract

Interest has been growing in decision-focused machine learning methods which train models to account for how their predictions are used in downstream optimization problems. Doing so can often improve performance on subsequent decision problems. However, current methods for uncertainty quantification do not incorporate any information about downstream decisions. We develop a methodology based on conformal prediction to identify prediction sets that account for a downstream cost function, making them more appropriate to inform high-stakes decision-making. Our approach harnesses the strengths of conformal methods -- modularity, model-agnosticism, and statistical coverage guarantees -- while incorporating downstream decisions and user-specified utility functions. We prove that our methods retain standard coverage guarantees. Empirical evaluation across a range of datasets and utility metrics demonstrates that our methods achieve significantly lower costs than standard conformal methods. We present a real-world use case in healthcare diagnosis, where our method effectively incorporates the hierarchical structure of dermatological diseases. The method successfully generates sets with coherent diagnostic meaning, potentially aiding triage for dermatology diagnosis and illustrating how our method can ground high-stakes decision-making employing domain knowledge.

Figures (8)

• Figure 1: Illustration of how proposed methods compare to the standard formulation of conformal prediction. The real-world example is taken from the Fitzpatrick dataset where the true disease is squamous cell carcinoma. The standard conformal prediction method results in a set of nine elements that span benign, malignant, and non-neoplastic diagnoses. Our method provides a formulation with a set of five elements within the malignant melanoma category.
• Figure 2: Empirical statistical coverage on the iNaturalist dataset across various loss functions. The observed coverage is close to the expected preset value of $\alpha$ for all 10 trials per loss function. We see a similar behavior for all the other datasets.
• Figure 3: Comparisons between different methods for the 1) separable loss case using the sum of penalties in each set as the loss metric (left), 2) non-separable loss case using maximum penalty in the set as the metric (middle), and 3) hierarchy coverage function loss case using the intersection of the set with each branch of the hierarchy as the metric (right). Numbers are the median-of-means across 10 different runs. See Appendix \ref{['appx:results']} for all datasets.
• Figure 4: Measured impact of base model accuracy on downstream optimization values on all losses considered for the cleaned Fitzpatrick dataset. Shows median-of-means across 10 different runs for each configuration. Results demonstrate that all methods improve with a higher quality model— i.e., a model with higher accuracy.
• Figure 5: Mean and standard deviation from the predicted softmax score for the true class as a function of predicted set size. The downward trend indicates that the predicted set size increases with the difficulty of the classification task for the model, highlighting the relationship between model uncertainty and set size.

Theorems & Definitions (9)

• Proposition 1
• Proposition 2
• Proposition 3
• Proposition 4
• proof
• proof
• proof
• proof
• proof