PreprintarXiv licencearXiv
Prior-data fitted networks for variable selection
Abstract
Prior-data fitted networks (PFNs) have become state-of-the-art for tabular prediction but remain underexplored in other areas of statistics. In this work, we present two PFNs for Bayesian variable selection: CD-PFN identifies conditional dependencies, and PC-PFN identifies parents and children in a causal graph. Both networks estimate posterior inclusion probabilities (PIPs) without posterior samples or data-specific modeling. In semi-synthetic and real data experiments, CD-PFN and PC-PFN often outperform existing methods, demonstrating the value of PFNs for variable selection and their potential in statistics more broadly. Open-source software is available.
Related
- Amortized Nonlinear Variable Selection with Prior-Data Fitted Networks
- SelectPFN: Amortized Bayesian Variable Selection for Sparse Nonlinear Regression with Prior-Data Fitted Networks
- Prior-Data Fitted Networks for In-Context Off-Policy Evaluation
- How Well Do Prior-Data Fitted Networks Amortize Bayesian Inference?
- Anisotropic Trust Region Warping for Prior-Data Fitted Networks in Bayesian Optimisation
BibTeX
@misc{melikechi2026priordata,
title = {{Prior-data fitted networks for variable selection}},
author = {Omar Melikechi},
howpublished = {arXiv},
year = {2026},
month = nov,
doi = {10.48550/arxiv.2611.05966},
eprint = {2611.05966},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2611.05966},
}
RIS
TY - UNPB
TI - Prior-data fitted networks for variable selection
AU - Omar Melikechi
PY - 2026
DA - 2026/11/04
PB - arXiv
DO - 10.48550/arxiv.2611.05966
UR - https://arxiv.org/abs/2611.05966
AB - Prior-data fitted networks (PFNs) have become state-of-the-art for tabular prediction but remain underexplored in other areas of statistics. In this work, we present two PFNs for Bayesian variable selection: CD-PFN identifies conditional dependencies, and PC-PFN identifies parents and children in a causal graph. Both networks estimate posterior inclusion probabilities (PIPs) without posterior samples or data-specific modeling. In semi-synthetic and real data experiments, CD-PFN and PC-PFN often outperform existing methods, demonstrating the value of PFNs for variable selection and their potential in statistics more broadly. Open-source software is available.
ER -