Anchor-Based Conformal Prediction Under Noisy Annotations in Single-Cell Data

Morgan He, Shijie Min, Grace Yi, Pingzhao Hu

Keywords: Conformal Prediction, Single-Cell Data, Noise Labels, Machine Learning

Abstract:

Conformal prediction provides a flexible framework for quantifying
prediction uncertainty and has attracted extensive interest. However, most
existing methods are designed to handle clean data and may fail to perform
satisfactorily when labels are noisy. In this work, we
consider the setting where the ground-truth labels are unobserved but crowdsourced noisy
labels are available. We introduce an anchor-based conformal prediction
method that provides uncertainty quantification.
Our method identifies anchor points by selecting samples
with strong agreement across annotators. These anchors points are used to train a base predictor
that is calibrated to construct a conformal prediction set with a desired coverage rate.
Meanwhile, we provide a theoretical analysis of anchor--point identification and
provide associated conditions that have been importantly overlooked in the literature.
We apply the proposed method to analyze two single-cell datasets to demonstrate its utility and promise.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 20743