Research map / Label-efficient and robust learning / Self-supervised and contrastive
SubTab: Subsetting Features of Tabular Data for Self-Supervised Representation Learning
Talip Ucar, Ehsan Hajiramezanali, Lindsay Edwards
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 |
|---|---|---|---|
| Reconstruct the full feature space from feature subsets rather than reconstructing only the input subset | SubTab · Tabular representation learning | downstream classification accuracy | Improved |
| Add swap-noise to the input (ablation) | SubTab · MNIST | accuracy: 97.26 -> 97.47 | Improved |
| Add distance loss (ablation) | SubTab with swap-noise · MNIST | accuracy: 97.47 -> 97.52 | Improved |
| Replace swap-noise with zero-out noise (ablation) | SubTab · MNIST | accuracy: 97.47 -> 97.25 | No clear effect |
| Replace swap-noise with Gaussian noise (ablation) | SubTab · MNIST | accuracy: 97.47 -> 97.25 | No clear effect |
| Add contrastive loss to reconstruction loss (ablation) | SubTab · MNIST | accuracy: 97.13 -> 97.26 | Improved |
| Use contrastive loss alone instead of reconstruction loss (ablation) | SubTab · MNIST | accuracy: 97.13 -> 97.11 | No clear effect |
| Shuffle feature order (ablation) | SubTab with swap-noise and distance loss · MNIST | accuracy: 97.52 -> 97.2 | No clear effect |
| Increase latent dimension from 128 to 512 while aggregating subset embeddings (ablation) | SubTab · MNIST | accuracy: 97.52 -> 97.86 | Improved |
| Concatenate subset embeddings instead of aggregating them (ablation) | SubTab with latent dimension 512 · MNIST | accuracy: 97.86 -> 95.92 | Hurt |
| Use a shallow one-layer encoder and decoder instead of the deep architecture (ablation) | SubTab · MNIST | accuracy: 97.86$\pm$0.07 -> 98.31$\pm$0.06 | Improved |
| Use a shallow one-layer encoder and decoder instead of the deep architecture (ablation) | SubTab · Obesity | accuracy: 71.13$\pm$4.08 -> 66.88$\pm$5.35 | Hurt |
Closest papers by meaning
- SwitchTab: Switched Autoencoders Are Effective Tabular Learners (AAAI 2024 · similarity 0.91)
- ConTextTab: A Semantics-Aware Tabular In-Context Learner (NeurIPS 2025 · similarity 0.90)
- Representation Space Augmentation for Effective Self-Supervised Learning on Tabular Data (AAAI 2025 · similarity 0.90)
- Large Scale Transfer Learning for Tabular Data via Language Modeling (NeurIPS 2024 · similarity 0.89)
- Hybrid Autoencoders for Tabular Data: Leveraging Model-Based Augmentation in Low-Label Settings (NeurIPS 2025 · similarity 0.88)
- ICE-T: Interactions-aware Cross-column Contrastive Embedding for Heterogeneous Tabular Datasets (AAAI 2025 · similarity 0.88)
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.
