Research map / Label-efficient and robust learning / Recognition and long-tail
Synthetic Tabular Data Generation for Imbalanced Classification: The Surprising Effectiveness of an Overlap Class
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 ORD, which relabels borderline majority examples as a third "overlap" class via random-forest k-fold disagreement, trains class-conditional generators (CTabSyn) on the ternary labels, and then trains classifiers on synthetic minority and clear-majority data only, to improve synthetic tabular data for imbalanced classification.
Problem
Deep tabular generators produce lower-quality minority examples than majority examples when trained on imbalanced data, which degrades classifiers trained on synthetic augmentation for imbalanced binary tabular classification.
Approach
Identify majority instances lying in the majority/minority overlap via random-forest k-fold disagreement, convert binary labels to ternary labels (clear majority, overlap majority, minority) for training a class-conditional generator, and then train the final classifier on synthetic minority and clear-majority data while discarding the overlap class.
Key results
- CTabSyn ORD achieves the best XGBoost accuracy on Adult (81.18 vs 80.01 for CTabSyn without ORD) and Cardio (73.10 vs 71.80), and ORD improves the average-of-4 accuracy for 26 of 32 (dataset, generator) combinations.
- A paired t-test over 5 synthesizers and 4 balanced datasets gives p=6.84e-05 for the ORD gain.
- On a 2D Gaussian-mixture setting with known Bayes labels, CTabSyn ORD raises minority generation accuracy to 96.50% versus 85.75% for TabSyn.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply ORD with CTabSyn rather than CTabSyn without ORD (ablation) | CTabSyn · Adult | Avg of 4 accuracy: 79.80 -> 80.67 | Helped |
| Apply ORD with CTabSyn rather than CTabSyn without ORD (ablation) | CTabSyn · Cardio | XGBoost accuracy: 71.80 -> 73.10 | Helped |
| Apply ORD with CTabSyn rather than CTabSyn without ORD (ablation) | CTabSyn · Fintech | Avg of 4 accuracy: 65.37 -> 63.58 | Hurt |
| Apply ORD with CTabSyn rather than CTabSyn without ORD (ablation) | CTabSyn · Heloc | XGBoost accuracy: 65.65 -> 67.77 | Helped |
| Add ternary overlap labels before generator training instead of filtering overlap after binary-label synthesis (ablation) | ORD synthesizer and classifier pipeline · Adult | XGBoost accuracy: 65.32 -> 79.19 | Helped |
| Add ternary overlap labels before generator training instead of filtering overlap after binary-label synthesis (ablation) | ORD synthesizer and classifier pipeline · Heloc | XGBoost accuracy: 44.72 -> 67.00 | Helped |
| Add real minority examples to synthetic ORD training data (ablation) | Synthetic ORD classifier training · Fintech | Avg of 4 accuracy: 60.52 -> 63.92 | Helped |
| Add overlap filtering by removing overlapping majority examples from real classifier training (ablation) | Classifier trained on real data · Adult | Accuracy: 76.23 -> 81.57 | Helped |
| Add overlap filtering by removing overlapping majority examples from real classifier training (ablation) | Classifier trained on real data · Heloc | Accuracy: 71.04 -> 70.55 | Hurt |
| Use ORD rather than the respective generator without ORD (ablation) | Five tabular synthesizers · Four balanced-test real datasets | Classification accuracy: 26 out of 32 | Mixed |
| Vary the overlap-detection threshold (ablation) | ORD with CTABGAN+ and ForestFlow · Adult | Avg of 4 accuracy: 74.63; 79.06; 78.69; 75.06; 72.30; 70.63 | Mixed |
| Vary the overlap-detection threshold (ablation) | ORD with CTABGAN+ and ForestFlow · Cardio | Avg of 4 accuracy: 72.16; 71.81; 71.01; 71.17; 71.37; 70.28 | Mixed |
Surprising findings
- Removing real majority instances from the classifier training pool (keeping only synthetic clear-majority and real minority) improved accuracy, despite generators being trained on the full data.
- Applying ORD only after synthesis (not before generator training) substantially reduced accuracy on Adult (65.32 vs 79.19) and Heloc (44.72 vs 67.00), indicating the ternary labels must be used during generator training.
Limitations
- Method is restricted to binary target columns; extension to multi-class is left open.
- Hyperparameters such as the overlap threshold tau and the ratio r are dataset- and generator-dependent and must be tuned on validation data.
- Authors note the method has not been tested on continuous skewed target columns.
Experimental setup
Datasets: Adult Heloc Fintech Cardio Abalone Bank Car Yeast
Benchmarks: Adult Heloc Fintech Cardio Abalone Bank Car Yeast 2D toy blobs
Metrics: macro-average accuracy on real minority and majority (machin minority accuracy majority accuracy F1 AUCROC
Baselines: SMOTE Borderline SMOTE ADASYN CTGAN CTAB-GAN+ TabDDPM ForestFlow TabSyn
Builds on: TabSyn SMOTE CTGAN ForestFlow random forest
Closest papers by meaning
- TabEBM: A Tabular Data Augmentation Method with Distinct Class-Specific Energy-Based Models (NeurIPS 2024 · similarity 0.88)
- EPIC: Effective Prompting for Imbalanced-Class Data Synthesis in Tabular Data Classification via Large Language Models (NeurIPS 2024 · similarity 0.88)
- CLIMB: Class-imbalanced Learning Benchmark on Tabular Data (NeurIPS 2025 · similarity 0.87)
- How Realistic Is Your Synthetic Data? Constraining Deep Generative Models for Tabular Data (ICLR 2024 · similarity 0.86)
- Balanced Data, Imbalanced Spectra: Unveiling Class Disparities with Spectral Imbalance (ICML 2024 · similarity 0.86)
- Generative Classifiers Avoid Shortcut Solutions (ICLR 2025 · similarity 0.86)
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.