Research map / Time series / Causal and generative time series
Time Weaver: A Conditional Time Series Generation Model
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 Time Weaver, a diffusion-based conditional time series generator that handles categorical, continuous and time-varying metadata, plus J-FTSD, a Fréchet-distance metric over joint time-series/metadata embeddings learned contrastively.
Problem
Generating realistic time series conditioned on heterogeneous paired metadata (categorical, continuous, time-varying), and evaluating whether generated series match metadata-specific features, which existing FD-based metrics ignore.
Approach
Adapt CSDI and SSSD diffusion denoisers with a metadata preprocessing module (one-hot/FC tokenizers plus self-attention) and introduce J-FTSD, which computes Fréchet distance over concatenated time-series and metadata embeddings trained jointly with CLIP-style contrastive loss.
Key results
- Time Weaver-CSDI reaches J-FTSD 2.2 vs 14.25 for WaveGAN on Air Quality and TSTR 0.77 vs 0.61.
- Time Weaver-SSSD achieves best ECG J-FTSD 5.43 vs WaveGAN 9.55 and TSTR 0.85 vs 0.65.
- Time Weaver is reported to beat the best GAN by about 6x on Air Quality, 1.75x on ECG, 4x on Electricity, and over 40x on Traffic in J-FTSD.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Time Weaver-CSDI instead of WaveGAN | Conditional time-series generation · Air Quality | J-FTSD: 14.25$\pm$0.79 -> 2.2$\pm$0.07 | Helped |
| Use Time Weaver-CSDI instead of WaveGAN | Conditional time-series generation · Air Quality | TSTR: 0.61$\pm$0.01 -> 0.77$\pm$0.01 | Helped |
| Use Time Weaver-SSSD instead of WaveGAN | Conditional time-series generation · ECG | J-FTSD: 9.55$\pm$0.01 -> 5.43$\pm$0.1 | Helped |
| Use Time Weaver-SSSD instead of WaveGAN | Conditional time-series generation · ECG | TSTR: 0.65$\pm$0.001 -> 0.85$\pm$0.007 | Helped |
| Use Time Weaver-SSSD instead of Pulse2Pulse | Conditional time-series generation · Traffic | J-FTSD: 17.70$\pm$0.002 -> 0.36$\pm$0.03 | Helped |
| Use Time Weaver-CSDI instead of WaveGAN | Conditional time-series generation · Traffic | TSTR: 0.54$\pm$0.01 -> 0.66$\pm$0.06 | Helped |
| Use Time Weaver-CSDI instead of Pulse2Pulse | Conditional time-series generation · Electricity | J-FTSD: 2.8$\pm$0.01 -> 0.6$\pm$0.003 | Helped |
| Use Time Weaver-CSDI instead of Pulse2Pulse | Conditional time-series generation · Electricity | TSTR: 0.71$\pm$0.004 -> 0.78$\pm$0.001 | Helped |
| Add paired metadata conditioning (ablation) | Time Weaver-CSDI and WaveGAN · Air Quality and Electricity | FTSD | Mixed |
| Train time-series and metadata feature extractors jointly rather than individually (ablation) | J-FTSD · Synthetic perturbations | J-FTSD sensitivity | Helped |
Surprising findings
- TimeGAN and TTS-GAN failed to converge on any dataset, even after 10,000 epochs.
- Without metadata, WaveGAN achieves lower FTSD than Time Weaver on Air Quality, though Time Weaver wins with metadata.
Limitations
- Diffusion models have slower inference and longer training than GAN baselines.
- Causal relationships between metadata and generated series are not rigorously established.
- Some baselines (TimeGAN, TTS-GAN) failed to converge, limiting comparison.
Experimental setup
Datasets: Electricity Traffic metro interstate volume Air Quality (Beijing) PTB-XL ECG
Benchmarks: Air Quality ECG Traffic Electricity
Metrics: J-FTSD TSTR (AUC) DTW FTSD
Baselines: WaveGAN Pulse2Pulse U-Net 1D diffusion
Builds on: CSDI SSSD CLIP Informer Context-FID
Closest papers by meaning
- Time Series Generation Under Data Scarcity: A Unified Generative Modeling Approach (NeurIPS 2025 · similarity 0.88)
- Generating Financial Time Series by Matching Random Convolutional Features (NeurIPS 2026 · similarity 0.86)
- TARFVAE: Efficient One-Step Generative Time Series Forecasting via TARFLOW based VAE (NeurIPS 2025 · similarity 0.86)
- TEMPO: Prompt-based Generative Pre-trained Transformer for Time Series Forecasting (ICLR 2024 · similarity 0.86)
- Sundial: A Family of Highly Capable Time Series Foundation Models (ICML 2025 · similarity 0.85)
- TS-RAG: Retrieval-Augmented Generation based Time Series Foundation Models are Stronger Zero-Shot Forecaster (NeurIPS 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.