为表格数据预测模型提供无需调参的不确定性量化方法
Uncertainty Quantification for Prior-Data Fitted Networks using Martingale Posteriors
- 基于鞅后验构建无须调参的采样推断流程
- 在模拟与真实数据上均实现良好校准与高效性
- 适合需要可靠置信区间的小中规模表格数据场景
先验-数据拟合网络（PFNs）作为表格数据预测的基石模型，无需调参即可在小到中等规模数据上达到顶尖性能。尽管其思想源于贝叶斯框架，但现有方法无法对预测均值、分位数等量提供不确定性量化。本文提出一种原则性强、高效且无需调参的采样方法，基于鞅后验构建此类估计的贝叶斯后验，并证明其收敛性。多个模拟与真实数据案例表明，该方法在推断应用中具备良好的效率与校准性。
原文摘要 · Abstract (English)
Prior-data fitted networks (PFNs) have emerged as promising foundation models for prediction from tabular datasets, achieving state-of-the-art performance on small to moderate data sizes without tuning. While PFNs are motivated by Bayesian ideas, they do not provide any uncertainty quantification for predictive means, quantiles, or similar quantities. We propose a principled, efficient, and tuning-free sampling procedure to construct Bayesian posteriors for such estimates based on martingale posteriors, and prove its convergence. Several simulated and real-world data examples showcase the efficiency and calibration of our method in inference applications.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。