Audited Conformal Prediction for Classification under Unknown Distribution Shift

2020-04-30 · Source: stat.ML updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

Audited Conformal Prediction (ACP) is a novel method for uncertainty quantification in classification models deployed under unknown distribution shifts. It addresses the degradation of predictive uncertainty estimates by integrating an auxiliary "audit" model, trained on a small labeled dataset from the target population, to identify inputs where the legacy model is likely to fail. ACP produces prediction sets that guarantee marginal coverage while significantly improving conditional coverage compared to existing approaches. The method offers two integration strategies: one for improved conditional performance with marginal coverage, and another for explicit group-conditional coverage guarantees. Experiments on synthetic data (e.g., K=5 classes, 100 features, 10,000 historical samples, 200-5,000 calibration samples) and real-world datasets like Camelyon17 and CIFAR-10/CIFAR-10-C demonstrate ACP's ability to balance reliability and efficiency, maintaining compact prediction set sizes while enhancing coverage for unreliable samples.

Key takeaway

For MLOps engineers deploying classification models in dynamic environments, Audited Conformal Prediction (ACP) offers a robust solution to maintain reliable uncertainty estimates under distribution shift. You should consider implementing ACP to improve conditional coverage for challenging samples without excessively inflating prediction set sizes. This approach allows you to adapt models efficiently with limited new labeled data, enhancing model trustworthiness and operational stability.

Key insights

Audited Conformal Prediction improves uncertainty quantification under distribution shift by using an auxiliary model to identify legacy model failures.

Principles

• An audit model can predict legacy model reliability without retraining the main model.
• Integrating audit model outputs into conformal prediction enhances conditional coverage.
• Data splitting is crucial for maintaining exchangeability and valid inference.

Method

Train a binary audit model îr on ij´ç²ç¹ to predict legacy model correctness. Integrate îr's output with the legacy model's predictions to construct conformal prediction sets on ij´ç²ç², ensuring marginal or group-conditional coverage.

In practice

• Use legacy model logits, confidence patterns, or intermediate features as audit model inputs.
• Employ random forests or neural networks for the audit model architecture.

Topics

• Conformal Prediction
• Distribution Shift
• Uncertainty Quantification
• Audit Models
• Classification Models
• Machine Learning Reliability

Code references

• FionaZ3696/Audited-Conformal-Prediction

Best for: Research Scientist, Computer Vision Engineer, AI Scientist, Machine Learning Engineer, MLOps Engineer

Related on AIssential

• Distribution-Free Uncertainty Quantification for Continuous AI Agent Evaluation — Conformal prediction offers robust, distribution-free uncertainty quantification for continuous AI agent evaluation.
• A Deployment Audit of Release-Side Risk in Conformal Triage under Prevalence Shift — A leakage-aware audit improves conformal triage safety by identifying event-positive releases under prevalence shift.
• Bridging Domain Expertise and Generalization for Performance Estimation — FRAP improves performance estimation under distribution shift by fusing foundation model robustness with base model domain expertise.
• Conformal Bayes under Label Shift: Post-Hoc Calibration vs. In-Training Adaptation — Conformal Bayes under label shift can be adapted via post-hoc calibration or in-training methods to maintain valid coverage.
• A Joint Finite-Sample Certificate for Adaptive Selective Conformal Risk Control — A new joint finite-sample certificate improves risk control and acceptance probability bounds for adaptive selective predictors.
• Conformal Blindness: A Note on $A$-Cryptic change-points — Conformal Test Martingales can be blind to significant data distribution shifts if p-values remain uniform.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by stat.ML updates on arXiv.org.