Problem
Research questions and friction points this paper is trying to address.
Tabular Foundation Models
Subpopulation Shift
Worst-group Accuracy
Distributionally Robust Adaptation
Parameter-Efficient Distributionally Robust Adaptation of Tabular Foundation Models under Subpopulation Shift
🤖 AI Summary
Tabular foundation models exhibit inadequate generalization to minority subgroups under subgroup distribution shifts. This work proposes the DR-TFM framework, which integrates distributionally robust optimization with parameter-efficient fine-tuning. By introducing an attention query scaling network that exclusively adapts the attention mechanism, DR-TFM achieves robust adaptation without requiring ground-truth labels. This approach overcomes the limitations of conventional full fine-tuning by updating merely 0.016% of model parameters. Experimental results across five benchmark datasets demonstrate that DR-TFM significantly improves worst-group accuracy while maintaining competitive average performance. Overall, this study presents an efficient solution for enhancing both fairness and robustness in tabular foundation models.
📝 Abstract
Despite strong mean accuracy, tabular foundation models (TFMs) can perform poorly on underrepresented groups under subpopulation shift, where group proportions change between training and deployment. We propose DR-TFM, a parameter-efficient distributionally robust adaptation framework that requires no true group annotations. DR-TFM adjusts attention to labeled context examples by fine-tuning an existing query scaling network or adding and training one, while keeping all other parameters fixed. We instantiate the framework with two robust objectives using estimated groups or source conditional distributions derived from training data. For TabPFN-3, adaptation updates only 0.016% of the pretrained model's parameters. Across five tabular benchmarks, DR-TFM achieves substantially higher average worst-group accuracy than pretrained TFMs and the compared robust baselines without true group annotations, while maintaining competitive mean group accuracy. DR-TFM also improves average worst-group accuracy on ACS Income and across four additional TFMs.
Innovation
Methods, ideas, or system contributions that make the work stand out.
Parameter-Efficient Adaptation
Distributionally Robust Optimization
Tabular Foundation Models
Subpopulation Shift
Worst-Group Accuracy
🔎 Similar Papers
No similar papers found.
