Research map / Label-efficient and robust learning / Self-supervised and contrastive
TransTab: Learning Transferable Tabular Transformers Across Tables
Zifeng Wang, Jimeng Sun
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TransTab instead of the strongest baseline | clinical-trial supervised learning · N00041119, N00174655, N00312208, N00079274, N00694382 | AUROC: 0.6397 -> 0.6408; 0.9310 -> 0.9428; 0.7586 -> 0.7770; 0.7178 -> 0.7281; 0.7458 -> 0.7648 | Improved |
| Use TransTab to learn from all three feature subsets instead of the strongest baseline | feature incremental learning · N00041119, N00174655, N00312208, N00079274, N00694382 | AUROC: 0.6371 -> 0.6797; 0.8485 -> 0.8545; 0.7031 -> 0.7617; 0.6535 -> 0.6857; 0.7236 -> 0.7795 | Improved |
| Pretrain across tables then finetune on the target | TransTab · clinical-trial mortality, across-table pretraining plus finetuning | AUROC: 0.6313 -> 0.6405; 0.8348 -> 0.8583; 0.7444 -> 0.7517; 0.6885 -> 0.7063; 0.7293 -> 0.7392 | Improved |
| Use vanilla supervised transfer learning rather than no pretraining (ablation) | TransTab · clinical-trial mortality, across-table pretraining plus finetuning | AUROC: 0.6313 -> 0.6424; 0.8348 -> 0.8183; 0.7444 -> 0.7458; 0.6885 -> 0.6928; 0.7293 -> 0.7239 | Mixed |
| Use Self-VPCL rather than no pretraining (ablation) | TransTab · public datasets, across-table pretraining plus finetuning | AUROC: 0.763 -> 0.777; 0.858 -> 0.837; 0.630 -> 0.626; 0.907 -> 0.907; 0.841 -> 0.819; 0.844 -> 0.843; 0.821 -> 0.823; 0.919 -> | Mixed |
| Use VPCL rather than no pretraining (ablation) | TransTab · public datasets, across-table pretraining plus finetuning | AUROC: 0.763 -> 0.776; 0.858 -> 0.858; 0.630 -> 0.637; 0.907 -> 0.907; 0.841 -> 0.862; 0.844 -> 0.844; 0.821 -> 0.819; 0.919 -> | Mixed |
| Use zero-shot learning from set1+set2 instead of supervised learning on set3 | TransTab · clinical-trial mortality zero-shot learning | AUROC: 0.5854 -> 0.5990; 0.6484 -> 0.6752; 0.7536 -> 0.7576; 0.7087 -> 0.7036; 0.6479 -> 0.6740 | Mixed |
| Use zero-shot learning from set1+set2 instead of supervised learning on set3 | TransTab · public datasets zero-shot learning | AUROC: 0.581 -> 0.685; 0.635 -> 0.721; 0.571 -> 0.538; 0.898 -> 0.892; 0.733 -> 0.710; 0.822 -> 0.804; 0.702 -> 0.742; 0.875 -> | Mixed |
| Pretrain on unrelated tables before finetuning (ablation) | TransTab · public datasets | AUROC: 0.763 -> 0.776; 0.858 -> 0.858; 0.630 -> 0.637; 0.907 -> 0.907; 0.841 -> 0.862; 0.844 -> 0.844; 0.821 -> 0.819; 0.919 -> | Mixed |
| Increase the number of vertical partitions (ablation) | VPCL versus Self-VPCL · clinical-trial and public datasets | AUROC | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- SCARF — Scarf: Self-Supervised Contrastive Learning using Random Feature Corruption (ICLR 2022) · better on 1 of 1 shared result
- VIME — VIME: Extending the Success of Self- and Semi-supervised Learning to Tabular Domain (NeurIPS 2020) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Large Scale Transfer Learning for Tabular Data via Language Modeling (NeurIPS 2024 · similarity 0.89)
- ConTextTab: A Semantics-Aware Tabular In-Context Learner (NeurIPS 2025 · similarity 0.87)
- CARTE: Pretraining and Transfer for Tabular Learning (ICML 2024 · similarity 0.86)
- Making Pre-trained Language Models Great on Tabular Prediction (ICLR 2024 · similarity 0.85)
- STRABLE: Benchmarking Tabular Machine Learning with Strings (NeurIPS 2026 · similarity 0.85)
- SynTab-LLaVA: Enhancing Multimodal Table Understanding with Decoupled Synthesis (CVPR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
