Distribution Shift Aware Neural Tabular Learning

Wangyang Ying, Nanxu Gong, Dongjie Wang, Xinyuan Wang, Arun Vignesh Malarkkan, Vivek Gupta, Chandan K. Reddy, Yanjie Fu · 2025

Read Distribution Shift Aware Neural Tabular Learning by Wangyang Ying, Nanxu Gong, Dongjie Wang, Xinyuan Wang, Arun Vignesh Malarkkan, Vivek Gupta, Chandan K. Reddy, Yanjie Fu (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-19486.txt.

Description

Eight researchers present a learning method that keeps table data models accurate when the data shifts, in a 2025 paper. Models trained on tables often fade when test data drifts away from training data. The team frames this as its own problem and proposes a framework called SAFT to fight it.

SAFT turns the search for good feature changes into a smooth, learnable generation process. Three guards keep it robust: decorrelated features with reweighted samples, averaged embeddings that favor flat, stable solutions, and normalization that aligns training and test data. Wide tests showed SAFT beat earlier table-learning methods under many kinds of shift. The gains held across varied real-world datasets. For teams whose data will not sit still, the method offers a sturdier base for prediction.

Catalog record

CREATORWangyang Ying, Nanxu Gong, Dongjie Wang, Xinyuan Wang, Arun Vignesh Malarkkan, Vivek Gupta, Chandan K. Reddy, Yanjie Fu
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE54 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-19486

Browse the full library · About the Tower of Babel