🤖 machine learning

Cost-Sensitive Conformal Prediction and Human-in-the-Loop Abstention for Imbalanced High-Stakes Decision Support: A Multi-Domain Benchmark

This paper presents a comprehensive multi-domain benchmark demonstrating that combining class-conditional (Mondrian) conformal prediction with cost-controlled abstention effectively resolves severe under-coverage of rare minority classes and minimizes expected decision costs in high-stakes, imbalanced decision support systems.

Original authors: Manpreet Singh, Akshatha Srikantha, Shyamal Lakhanpal

Published 2026-07-30

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Manpreet Singh, Akshatha Srikantha, Shyamal Lakhanpal

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are the captain of a spaceship navigating a dense, foggy asteroid field. Your ship's computer is brilliant at spotting the thousands of harmless, floating rocks (the "majority" class), but it keeps missing the rare, dangerous asteroids that could destroy the ship (the "minority" class). In the world of artificial intelligence, this is a common problem called "imbalanced data." When a computer learns from a dataset where one outcome happens 99 times and another happens only once, it gets lazy and starts guessing the common outcome every time to look smart. This is dangerous in high-stakes jobs like spotting credit card fraud, diagnosing rare diseases, or predicting factory machine failures, where missing the rare event costs a fortune or even lives.

To fix this, scientists use a safety net called "Conformal Prediction." Think of this as a smart guardrail that doesn't just give a single "yes" or "no" answer, but instead says, "I am 90% sure the answer is in this group." If the computer is confident, the group has one item. If it's confused, the group might have two items, or even be empty, signaling, "I have no idea, please ask a human!" The big question researchers have been asking is: Does this safety net work equally well for the rare, dangerous asteroids, or does it just protect the boring, common rocks?

This paper, titled "Cost-Sensitive Conformal Prediction and Human-in-the-Loop Abstention for Imbalanced High-Stakes Decision Support," dives deep into that exact problem. The authors, Manpreet Singh, Akshatha Srikantha, and Shyamal Lakhanpal, ran a massive experiment across 15 different real-world datasets, testing 7 different types of AI models. They discovered that the standard version of this safety net (called "Marginal Conformal Prediction") is actually failing the rare cases spectacularly. In their tests, while the system promised 90% safety overall, it was only catching the rare, dangerous events less than 1% of the time in the worst cases. It was like a guardrail that protects the sidewalk but leaves the cliff edge completely open.

The paper's main finding is that a specific upgrade called "Mondrian Conformal Prediction" fixes this broken guardrail. Instead of using one giant safety net for everything, Mondrian CP builds separate, custom nets for the common rocks and the rare asteroids. The results were dramatic: this method restored the safety coverage for the rare events to a solid 92.2% on average, a massive improvement of 61.7 percentage points over the standard method.

But the paper doesn't stop at just catching the rare events; it also figures out when it's worth paying a human to step in. The authors introduced a "cost-controlled abstention" rule. Imagine the AI is a student taking a test. If the student is unsure, they can choose to "abstain" (leave the answer blank) and let a teacher grade it, but the teacher charges a fee for their time. The paper calculated the exact "break-even" point: if the fee for the teacher is low enough, it is mathematically cheaper to let the AI skip the hard questions and send them to a human, rather than risking a wrong answer that costs a fortune. They proved that by using their improved Mondrian method combined with this smart "skip and ask" rule, the total cost of making decisions dropped significantly compared to standard methods.

In short, the paper argues that for high-stakes decisions involving rare events, the old "one-size-fits-all" safety nets are dangerous. Instead, we need custom nets for rare cases and a smart system that knows exactly when to say, "I don't know, let a human handle this," to save money and prevent disasters. The authors tested this across thousands of scenarios and found that this approach is not just a good idea, but a statistically proven necessity for keeping rare, costly errors in check.

Technical Summary: Cost-Sensitive Conformal Prediction and Human-in-the-Loop Abstention for Imbalanced High-Stakes Decision Support

Problem Statement
High-stakes decision systems in domains such as credit scoring, fraud detection, healthcare, and industrial safety operate under severe class imbalance and asymmetric error costs. In these contexts, the cost of a false negative (e.g., missing a fraudulent transaction or a severe disease) often vastly exceeds the cost of a false positive. While standard classification models provide point predictions or confidence scores, they frequently fail to quantify true prediction uncertainty, leading to risky automated decisions.

Conformal Prediction (CP) offers a distribution-free framework for uncertainty quantification, guaranteeing that the true label lies within a prediction set with a probability of at least $1-\alpha$ . However, standard marginal conformal prediction provides validity only averaged across the entire data distribution. On imbalanced datasets, this results in a critical failure mode: the predictor satisfies the global coverage target by over-covering the majority class while severely under-covering the rare, high-cost minority class. Empirical evidence in this study shows that marginal CP can drop minority class coverage to below 1% on extreme imbalance ratios, leaving rare but critical cases unprotected.

Methodology
The authors propose a comprehensive framework integrating Class-Conditional (Mondrian) Conformal Prediction with Cost-Controlled Abstention.

Mondrian Conformal Prediction: To address the minority coverage deficit, the method splits the calibration set into class-specific subsets ( $D^{(0)}_{cal}$ and $D^{(1)}_{cal}$ ). Independent non-conformity quantiles ( $q_0$ and $q_1$ ) are computed for each class. This ensures that the coverage guarantee $P(Y \in \hat{C}(X) \mid Y=y) \geq 1-\alpha$ holds for every class individually, rather than just marginally.
Cost-Controlled Abstention Mechanism: The system maps prediction set outputs to actions based on an asymmetric cost matrix ( $C_{FN}, C_{FP}$ $C_{F N}, C_{F P}$ ) and a human review cost ( $C_{rev}$ $C_{r e v}$ ).
- If the prediction set is a singleton (e.g., $\{0\}$ or $\{1\}$ ), the system executes the automated decision.
- If the prediction set is ambiguous (multi-label $\{0, 1\}$ ) or empty ( $\emptyset$ ), the system abstains and defers the instance to a human expert.
- The framework models the expected cost of deferral, accounting for the operational cost of review and potential human error rates ( $\epsilon_{hum}$ ).
Break-Even Analysis: The authors derive a closed-form threshold ( $C^*_{rev}$ ) representing the maximum review cost at which deferring ambiguous instances to humans yields a lower total expected decision cost compared to standard point predictors or Bayes-optimal threshold tuning.

Experimental Benchmark
The study conducts a rigorous empirical evaluation across:

15 Real-World Datasets: Sourced from OpenML, covering finance, healthcare, industrial safety, and environmental monitoring, with imbalance ratios ranging from 2.3:1 to 580:1.
7 Base Classifiers: Including HistGradientBoosting, Logistic Regression, Random Forest, Extra Trees, Gradient Boosting, AdaBoost, and Gaussian Naive Bayes.
3 Calibration Techniques: Uncalibrated, Platt Scaling, and Isotonic Regression.
Scale: 3,150 total experimental runs (10 random seeds per configuration).

Key Results

Restoration of Minority Coverage: Standard marginal CP achieved an average minority class coverage of only 30.5%, often dropping below 1% on extreme datasets (e.g., aps_failure). In contrast, Mondrian CP systematically restored valid coverage, achieving an average minority coverage of 92.2%. This represents an average improvement of 61.7 percentage points over marginal CP ( $p < 10^{-80}$ ).
Cost Reduction: Coupling Mondrian CP with cost-controlled abstention significantly reduced the expected per-instance decision cost. At a moderate human review cost ( $C_{rev} = 0.5$ ), the proposed method achieved a mean cost of 0.313, a 54.1% reduction compared to default point classification and a 38.6% reduction compared to Bayes cost-tuned thresholds.
Break-Even Thresholds: The study quantified dataset-specific break-even thresholds for human review. It demonstrated that even with noisy human review ( $\epsilon_{hum} > 0$ ), deferral remains economically advantageous provided the human error rate remains below approximately 18% under standard operating parameters.
Robustness: The coverage restoration and cost benefits were consistent across all seven base classifiers and three calibration methods, confirming the model-agnostic nature of the approach.

Significance and Claims
The paper claims to provide actionable principles for deploying distribution-free, cost-optimal uncertainty quantification in knowledge-based decision support systems. Its primary contributions are:

Empirical Quantification: Demonstrating that marginal CP is fundamentally inadequate for imbalanced high-stakes tasks due to severe minority under-coverage.
Systematic Solution: Proving that Mondrian CP restores validity for rare classes without requiring complex model retraining.
Economic Framework: Establishing a formal link between conformal set-valued outputs and financial/clinical decision costs, deriving specific thresholds where human-in-the-loop deferral is economically superior to automated tuning.
Deployment Guidelines: Offering a four-tier system architecture for integrating these methods into enterprise environments, including considerations for human capacity limits (queuing theory) and distribution shifts.

The authors conclude that while the method relies on tabular data and assumes exchangeability, it provides a robust, theoretically grounded approach to managing uncertainty and cost in critical decision-making scenarios where rare errors carry disproportionate consequences. Future work is suggested in online adaptive inference, LLM-assisted routing, fairness-aware extensions, and multi-class generalization.

Technical Summary: Cost-Sensitive Conformal Prediction and Human-in-the-Loop Abstention for Imbalanced High-Stakes Decision Support

More like this