DANCE: Doubly Adaptive Neighborhood Conformal Estimation

TL;DR

DANCE introduces a doubly adaptive conformal prediction method that leverages data embeddings and kernel regression to produce efficient and robust uncertainty sets for classification, outperforming existing approaches.

Contribution

It proposes a novel doubly locally adaptive conformal algorithm combining two nonconformity scores using embedded representations, enhancing set efficiency and robustness.

Findings

DANCE outperforms state-of-the-art conformal methods in set size efficiency.

DANCE demonstrates robustness across various datasets.

The method effectively balances set size and confidence.

Abstract

The recent developments of complex deep learning models have led to unprecedented ability to accurately predict across multiple data representation types. Conformal prediction for uncertainty quantification of these models has risen in popularity, providing adaptive, statistically-valid prediction sets. For classification tasks, conformal methods have typically focused on utilizing logit scores. For pre-trained models, however, this can result in inefficient, overly conservative set sizes when not calibrated towards the target task. We propose DANCE, a doubly locally adaptive nearest-neighbor based conformal algorithm combining two novel nonconformity scores directly using the data's embedded representation. DANCE first fits a task-adaptive kernel regression model from the embedding layer before using the learned kernel space to produce the final prediction sets for uncertainty quantification. We test against state-of-the-art local, task-adapted and zero-shot conformal baselines, demonstrating DANCE's superior blend of set size efficiency and robustness across various datasets.