Research map / Label-efficient and robust learning / Semi-supervised learning
Geodesic Flow Kernels for Semi-Supervised Learning on Mixed-Variable Tabular Dataset
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes GFTab, a semi-supervised framework for mixed-variable tabular data combining variable-type-specific corruption, a geodesic flow kernel similarity loss on the Grassmannian, and tree-based (GBDT leaf) embeddings, evaluated on 21 curated tabular datasets.
Problem
Semi-supervised learning on tabular data with mixed continuous and categorical variables and limited labels; existing corruption-based methods ignore variable types and are evaluated mostly on continuous-only data.
Approach
Uses different corruption schemes for continuous (soft/hard permutation mixing after a GRN-based variable selection network) and categorical variables (neighborhood-shift corruption), and measures similarity between soft and hard views via a geodesic flow kernel on the Grassmannian, plus GBDT leaf embeddings learned from labeled data.
Key results
- In the 20% labeled setting, GFTab achieved the top F1 on 10 of 21 datasets and top-three on 7 more (Table 1).
- With 20% label noise, GFTab had the highest F1 on 8 datasets and top-three on 10 others.
- Ablation shows removing tree-based embedding lowers accuracy on several datasets, e.g., credit-approval 0.7853 vs 0.7206 and qsar 0.7099 vs 0.6472 (Appendix D).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add variable selection network (VSN) (ablation) | GFTab · Nine tabular datasets, 20% label noise | F1 score: 8.770; 2.037; 4.201; 1.412; 0.877; 3.307; 3.344; -4.017; 4.947 | Mixed |
| Add tree-based embedding (ablation) | GFTab · Nine tabular datasets, 20% label noise | F1 score: 6.699; 8.239; -0.635; 5.369; 4.806; -0.241; 0.733; 4.982; 8.832 | Mixed |
| Use geodesic flow kernel loss instead of Uniform-Alignment, Barlow Twins, or InfoNCE (ablation) | GFTab · 21 benchmark datasets without label noise | win count: at least 14 out of 21 datasets | Helped |
| Set balance parameter β to 1.0 (ablation) | GFTab · 10% labeled, 10% label noise | F1 score | No clear effect |
| Set balance parameter β to 0.4 instead of 1.0 (ablation) | GFTab · 10% labeled, 10% label noise | F1 score | Helped |
| Use GFTab categorical corruption instead of Permute or Random (ablation) | GFTab · Benchmark datasets, including noisy-label settings | win count | Helped |
Surprising findings
- GFTab did not lose on dresses-sales and shrutime when tree embedding was removed (slight decreases of -0.64% and -0.24%).
Limitations
- Results are not universally superior; some baselines (e.g., XGBoost, CatBoost, SubTab) win on several datasets.
- Dataset curation criteria (i.i.d., mixed types, size bins) limit generality; around 60% of datasets excluded.
- Hyperparameter search for GFTab uses a limited space; single-GPU experiments with modest search budgets.
Experimental setup
Datasets: cmc jasmine credit-g Online_shoppers dresses-sales shrutime credit-approval Insurance bank fars okcupid-stem adult
Benchmarks: 21 curated tabular datasets (OpenML/AMLB-derived)
Metrics: F1 score win matrix
Baselines: GRANDE TabPFN SCARF VIME SubTab XGBoost CatBoost
Builds on: VIME SCARF SubTab DATE GRANDE
Closest papers by meaning
- TRL-Bench: Standardizing Cross-Paradigm Representation-Level Evaluation of Tabular Encoders (NeurIPS 2026 · similarity 0.85)
- TabGLM: Tabular Graph Language Model for Learning Transferable Representations Through Multi-Modal Consistency Minimization (AAAI 2025 · similarity 0.84)
- A Data-Centric Perspective on Evaluating Machine Learning Models for Tabular Data (NeurIPS 2024 · similarity 0.84)
- Hybrid Autoencoders for Tabular Data: Leveraging Model-Based Augmentation in Low-Label Settings (NeurIPS 2025 · similarity 0.84)
- SwitchTab: Switched Autoencoders Are Effective Tabular Learners (AAAI 2024 · similarity 0.84)
- Representation Space Augmentation for Effective Self-Supervised Learning on Tabular Data (AAAI 2025 · similarity 0.84)
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.