Conformal Prediction for Tabular Prior-Data Fitted Networks with Missing data

Florian D. van Leeuwen

Submission Type: Short paper (4 pages)

Keywords: TabPFN, missing values, conformal prediction, predictive uncertainty quantification

TL;DR: TabPFN produces invalid uncertainty estimates when data has missing values, but adapting conformal prediction (CP-MDA-exact) fixes this issue to provide mask-conditional-validity.

Abstract:

Tabular Prior-Data Fitted Networks (PFNs) achieve state-of-the-art performance for prediction tasks on small tabular datasets. Beyond point predictions, PFNs provide full posterior predictive distributions for uncertainty quantification. Among these models, TabPFN has emerged as particularly powerful due to its ability to handle missing data through internal deterministic imputation. However, we demonstrate that its uncertainty estimates are not valid conditional on the missing data pattern.

In this paper, we adapt the conformal prediction method CP-MDA-exact to TabPFN, providing a practical framework for obtaining mask-conditional valid uncertainty estimates. Our experiments on simulated data demonstrate that this approach successfully corrects the coverage across different missing patterns, even with small calibration sets.

Submission Number: 29