Research map / Time series / Causal and generative time series
TSGM: A Flexible Framework for Generative Modeling of Synthetic Time Series
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.
Introduces TSGM, an open-source Keras-based framework that unifies data-driven (GAN, VAE, probabilistic) and simulation-based generators of synthetic time series with a broad suite of evaluation metrics and augmentations.
Problem
Synthetic time series generation lacks a unified software framework and common metrics for comparing methods, limiting data sharing and augmentation for scarce or sensitive temporal data.
Approach
A single extensible API (built on Keras 3) that combines data-driven generators (GAN, VAE, TimeGAN, RCGAN, TimeVAE, beta-VAE), simulation-based generators with ABC rejection sampling, augmentations, and multi-dimensional evaluation metrics (similarity, predictive consistency, privacy, fairness, downstream effectiveness, diversity, visual comparison) for two scenarios: data sharing and data augment
Key results
- On ECG-200 with a 1D-CNN classifier, WindowWarp augmentation gave the highest accuracy and F1 (0.91 ± 0.01) versus 0.89 ± 0.02 for the no-augmentation baseline.
- Hardware benchmark training a cGAN on 5000 univariate series for two epochs: CPU ~25m39s per first two epochs, single V100 GPU ~24.9s, two GPUs ~34.4s, TPU v2 ~1m34s.
- Demo 1 compares TimeVAE, TimeGAN and RCGAN across similarity, consistency, downstream, and privacy metrics on C-MAPSS, periodic, and UCI energy data, reported as heatmaps (specific values not given in the text).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add GaussianNoise augmentation | CNN1d trained on ECG-200 with 100 additional samples · ECG-200 | accuracy: 0.89 -> 0.89 | No clear effect |
| Add GaussianNoise augmentation | CNN1d trained on ECG-200 with 100 additional samples · ECG-200 | F1: 0.89 -> 0.89 | No clear effect |
| Add SliceAndShuffle augmentation | CNN1d trained on ECG-200 with 100 additional samples · ECG-200 | accuracy: 0.89 -> 0.89 | No clear effect |
| Add SliceAndShuffle augmentation | CNN1d trained on ECG-200 with 100 additional samples · ECG-200 | F1: 0.89 -> 0.89 | No clear effect |
| Add Rotation augmentation | CNN1d trained on ECG-200 with 100 additional samples · ECG-200 | accuracy: 0.89 -> 0.90 | No clear effect |
| Add Rotation augmentation | CNN1d trained on ECG-200 with 100 additional samples · ECG-200 | F1: 0.89 -> 0.90 | No clear effect |
| Add MagnitudeWarp augmentation | CNN1d trained on ECG-200 with 100 additional samples · ECG-200 | accuracy: 0.89 -> 0.90 | No clear effect |
| Add MagnitudeWarp augmentation | CNN1d trained on ECG-200 with 100 additional samples · ECG-200 | F1: 0.89 -> 0.90 | No clear effect |
| Add WindowWarp augmentation | CNN1d trained on ECG-200 with 100 additional samples · ECG-200 | accuracy: 0.89 -> 0.91 | Helped |
| Add WindowWarp augmentation | CNN1d trained on ECG-200 with 100 additional samples · ECG-200 | F1: 0.89 -> 0.91 | Helped |
| Add DTWBA augmentation | CNN1d trained on ECG-200 with 100 additional samples · ECG-200 | accuracy: 0.89 -> 0.90 | No clear effect |
| Add DTWBA augmentation | CNN1d trained on ECG-200 with 100 additional samples · ECG-200 | F1: 0.89 -> 0.90 | No clear effect |
Surprising findings
- Demo 1 indicates that data sharing and data augmentation scenarios require different generative modeling choices, so no single method dominates across metrics.
Limitations
- Not tested on irregularly sampled time series and lacks state-of-the-art models for that setting.
- Empirical demonstrations are small-scale and mostly qualitative; results are shown as heatmaps and plots without full numerical tables.
Experimental setup
Datasets: NASA C-MAPSS UCI Appliances energy ECG-200 UCR Physionet 2012 MNIST Mauna Loa CO2 EEG eye state Individual Household Electric Power Consumption COVID-19 US Gaussian process samples
Benchmarks: NASA C-MAPSS UCI Appliances energy periodic/sine-const synthetic data ECG-200
Metrics: summary statistic distance MMD discriminative score predictive consistency membership inference privacy downstream MSE improvement classification accuracy F1 t-SNE training time
Baselines: TimeSynth DeepEcho Synthcity TSSurrogates Gretel TimeVAE TimeGAN RCGAN
Builds on: TimeGAN TimeVAE beta-VAE RCGAN Wasserstein GAN
Closest papers by meaning
- Time Series Generation Under Data Scarcity: A Unified Generative Modeling Approach (NeurIPS 2025 · similarity 0.90)
- TimeTok: Granularity-Controllable Time-Series Generation via Hierarchical Tokenization (NeurIPS 2026 · similarity 0.87)
- Synthetic Series-Symbol Data Generation for Time Series Foundation Models (NeurIPS 2025 · similarity 0.86)
- HDT: Hierarchical Discrete Transformer for Multivariate Time Series Forecasting (AAAI 2025 · similarity 0.86)
- TEMPO: Prompt-based Generative Pre-trained Transformer for Time Series Forecasting (ICLR 2024 · similarity 0.86)
- TimeDP: Learning to Generate Multi-Domain Time Series with Domain Prompts (AAAI 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.