Research map / Learning theory and algorithms / Prediction and decision losses
U-trustworthy Models. Reliability, Competence, and Confidence in Decision-Making
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes U-trustworthiness, a competence-based definition of trustworthiness for binary classifiers in utility-maximizing decision tasks, showing properly-ranked classifiers are U-trustworthy and advocating AUC over calibration/accuracy for model selection and tuning.
Problem
Existing trustworthy AI evaluation relies on probabilistic framing and calibration as prerequisites for trustworthiness, which may mislead model selection for utility-maximizing decision tasks.
Approach
Defines U-trustworthiness as a model achieving maximum Bayes-level expected utility over a class of utility functions; shows properly-ranked classifiers (order-preserving transforms of the Bayes classifier) are U-trustworthy regardless of calibration, and that AUC equals Bayes AUC iff the classifier is properly ranked, so AUC is a principled trustworthiness measure.
Key results
- On the homeownership task, Random Forest has the highest AUC (0.797) while Logistic Regression has the highest accuracy (0.736) and lowest Brier (0.177); the AUC-selected model achieves the highest maximum utility (0.744).
- In k-NN tuning, selecting k=200 by AUC yields higher test utility than k=150 selected by accuracy across utility thresholds.
- On Bankruptcy, Random Forest (AUC 0.928) is preferred by AUC while NetTrust would select LR, showing NetTrust and accuracy-based selection can disagree with maximum utility.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Select Random Forest by AUC instead of Logistic Regression by accuracy or Brier score | Homeownership model selection · Homeownership | Maximum utility: 0.739 -> 0.744 | Helped |
| Select Random Forest by AUC instead of kNN by NetTrust | Bankruptcy model selection · Bankruptcy | Maximum utility: 0.968 -> 0.971 | Helped |
| Tune k using AUC instead of accuracy (k=200 instead of k=150) | kNN · Homeownership | Test utility | Helped |
| Use a calibrated classifier rather than a properly-ranked classifier | Classifier comparison · Simulated classification data | Maximum utility | Hurt |
| Use a properly-ranked classifier rather than a calibrated classifier | Classifier comparison · Simulated classification data | Maximum utility | Helped |
| Select by accuracy instead of AUC | kNN hyperparameter tuning · Homeownership | Test utility | Hurt |
Surprising findings
- A calibrated classifier can be U-incompetent, while a miscalibrated properly-ranked classifier attains Bayes-level maximum utility.
- NetTrust score ranks the Bayes classifier below the properly-ranked classifier, suggesting it can mislead trustworthiness assessments.
Limitations
- Theory is restricted to binary classifiers and does not extend to multi-class settings.
- Applies only to utility-maximization tasks with decision-rule solutions of threshold form; not to risk mitigation or population inference.
- The AUC of a U-trustworthy model is unknown, making confidence via hypothesis testing difficult.
Experimental setup
Datasets: 2019 American Housing Survey homeownership data (48,660 households) Adult (UCI) Bankruptcy (UCI Taiwanese companies) Breast Cancer Wisconsin (UCI, 569 instances) simulated data (400 datasets, 15,000 samples each)
Benchmarks: homeownership (AHS 2019) Adult Bankruptcy Breast Cancer
Metrics: AUC accuracy Brier score NetTrust score maximum expected utility
Baselines: Brier score accuracy NetTrust score calibration error
Builds on: Bayes classifier competence-based trust theory NetTrust score
Closest papers by meaning
- Trustworthy AI Must Account for Interactions (NeurIPS 2026 · similarity 0.89)
- To trust or not to trust: Attention-based Trust Management for LLM Multi-Agent Systems (NeurIPS 2026 · similarity 0.87)
- More RLHF, More Trust? On The Impact of Preference Alignment On Trustworthiness (ICLR 2025 · similarity 0.87)
- Position: TrustLLM: Trustworthiness in Large Language Models (ICML 2024 · similarity 0.85)
- Learning to Decide with AI Assistance under Human-Alignment (NeurIPS 2026 · similarity 0.85)
- Trustworthy Machine Learning through Data-Specific Indistinguishability (ICML 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.