29
15

Conformal prediction under ambiguous ground truth

Abstract

Conformal Prediction (CP) allows to perform rigorous uncertainty quantification by constructing a prediction set C(X)C(X) satisfying P(YC(X))1α\mathbb{P}(Y \in C(X))\geq 1-\alpha for a user-chosen α[0,1]\alpha \in [0,1] by relying on calibration data (X1,Y1),...,(Xn,Yn)(X_1,Y_1),...,(X_n,Y_n) from P=PXPYX\mathbb{P}=\mathbb{P}^{X} \otimes \mathbb{P}^{Y|X}. It is typically implicitly assumed that PYX\mathbb{P}^{Y|X} is the "true" posterior label distribution. However, in many real-world scenarios, the labels Y1,...,YnY_1,...,Y_n are obtained by aggregating expert opinions using a voting procedure, resulting in a one-hot distribution PvoteYX\mathbb{P}_{vote}^{Y|X}. For such ``voted'' labels, CP guarantees are thus w.r.t. Pvote=PXPvoteYX\mathbb{P}_{vote}=\mathbb{P}^X \otimes \mathbb{P}_{vote}^{Y|X} rather than the true distribution P\mathbb{P}. In cases with unambiguous ground truth labels, the distinction between Pvote\mathbb{P}_{vote} and P\mathbb{P} is irrelevant. However, when experts do not agree because of ambiguous labels, approximating PYX\mathbb{P}^{Y|X} with a one-hot distribution PvoteYX\mathbb{P}_{vote}^{Y|X} ignores this uncertainty. In this paper, we propose to leverage expert opinions to approximate PYX\mathbb{P}^{Y|X} using a non-degenerate distribution PaggYX\mathbb{P}_{agg}^{Y|X}. We develop Monte Carlo CP procedures which provide guarantees w.r.t. Pagg=PXPaggYX\mathbb{P}_{agg}=\mathbb{P}^X \otimes \mathbb{P}_{agg}^{Y|X} by sampling multiple synthetic pseudo-labels from PaggYX\mathbb{P}_{agg}^{Y|X} for each calibration example X1,...,XnX_1,...,X_n. In a case study of skin condition classification with significant disagreement among expert annotators, we show that applying CP w.r.t. Pvote\mathbb{P}_{vote} under-covers expert annotations: calibrated for 72%72\% coverage, it falls short by on average 10%10\%; our Monte Carlo CP closes this gap both empirically and theoretically.

View on arXiv
Comments on this paper