Research map / Time series / Causal and generative time series
Diffusion Transformers for Tabular Data Time Series Generation
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 TabDiT, a latent diffusion transformer that generates heterogeneous tabular data time series by combining per-row VAE embeddings, with an autoregressive VAE decoder, variable-range decimal number encoding, and explicit end-of-sequence padding prediction.
Problem
Generating time series of heterogeneous tabular data (mixed numerical and categorical fields with variable length) remains largely unexplored; existing autoregressive methods lack diverse unconditional sampling and single-row diffusion methods assume fixed-length inputs.
Approach
Use a non-holistic VAE that encodes individual tabular rows, and a DiT denoiser that combines the row embeddings into time series; add an AR VAE decoder, a magnitude-prefix decimal representation for numbers, and generate the sequence length by predicting EoS padding rows.
Key results
- Unconditional generation on Age1: TabDiT MLD-TS 63.93 vs AR Baseline 91.20; on Age2 TabDiT reaches MLD-TS 50.43 (near chance).
- Conditional 'child gt-cond' on Rossmann: TabDiT MLD-TS 72.20 and LD-SR 82.90, versus REaLTabFormer 98.90 and 60.63.
- Ablation on Age2: full TabDiT MLD-TS 50.8, versus 62.2 for parallel VAE decoder and 83.7 for linear transformation numeric encoding.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace the AR Baseline with TabDiT | Unconditional time-series generation · Rossmann, Airbnb, PKDD’99 Financial, Age2, Age1, Leaving | MLD-TS: 97.80 -> 82.60; 77.23 -> 55.07; 92.87 -> 85.53; 67.53 -> 50.43; 91.20 -> 63.93; 69.43 -> 62.33 | Helped |
| replace the AR Baseline with TabDiT | Conditional generation, child gt-cond · Rossmann, Airbnb, Age2, PKDD’99 Financial | MLD-TS: 95.57 -> 72.20; 57.97 -> 51.10; 69.97 -> 51.40; 68.33 -> 59.50 | Helped |
| add an autoregressive VAE decoder instead of a parallel decoder (ablation) | TabDiT · Age2 | MLD-TS: 62.2 -> 50.8 | Helped |
| add decoder cross-attention to the latent z (ablation) | TabDiT · Age2 | MLD-TS: 51.2 -> 50.8 | No clear effect |
| use variable-range decimal representation instead of fixed-length digits (ablation) | TabDiT · Age2 | MLD-TS: 64.0 -> 50.8 | Helped |
| use variable-range decimal representation instead of quantized numerical features (ablation) | TabDiT · Age2 | MLD-TS: 83.8 -> 50.8 | Helped |
| use variable-range decimal representation instead of a linear transformation for numerical values (ablation) | TabDiT · Age2 | MLD-TS: 83.7 -> 50.8 | Helped |
| predict sequence length explicitly instead of ignoring padding and sampling length from a Gaussian (ablation) | TabDiT · Age2 | MLD-TS: 61.0 -> 50.8 | Helped |
| use the diffusion model rather than the autoregressive baseline | TabDiT · Age2 | MLD-TS: 68.3 -> 50.8 | Helped |
Surprising findings
- On the Age2 dataset, TabDiT even surpasses the ideal 'Original' MLE-TS accuracy (61.57 vs 60.07).
Limitations
- Single-row generation is not optimized; the single-row variant underperforms REaLTabFormer on parent generation in most datasets.
- Large-scale experiments use a private dataset that cannot be released.
- Unconditional comparison only against the authors' own AR baseline since no public unconditional baseline exists.
Experimental setup
Datasets: Age1 Age2 Leaving PKDD'99 Financial Rossmann Airbnb Large Scale Bank Data
Benchmarks: Age1 Age2 Leaving PKDD'99 Financial Rossmann Airbnb
Metrics: MLD-TS LD-SR MLE-TS
Baselines: REaLTabFormer SDV AR Baseline TabGPT-style quantization TabSyn-style linear transform
Builds on: DiT Latent Diffusion Models TabSyn FiT REaLTabFormer
Closest papers by meaning
- TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation (ICLR 2025 · similarity 0.88)
- Continuous Diffusion for Mixed-Type Tabular Data (ICLR 2025 · similarity 0.87)
- Utilizing Image Transforms and Diffusion Models for Generative Modeling of Short and Long Time Series (NeurIPS 2024 · similarity 0.87)
- TabNAT: A Continuous-Discrete Joint Generative Framework for Tabular Data (ICML 2025 · similarity 0.86)
- Time Series Generation Under Data Scarcity: A Unified Generative Modeling Approach (NeurIPS 2025 · similarity 0.86)
- A Diffusion Model for Regular Time Series Generation from Irregular Data with Completion and Masking (NeurIPS 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.