back to list

Project: High-Quality Long Time Series Generation with Deep Generative Models: Exploring Flow Matching and Exogenous Data Integration (TAKEN)

Description

Time series data—prevalent in finance, healthcare, and climate science—requires advanced generative models for tasks like data augmentation, anomaly detection, scenario planning, and synthetic data generation. Deep Generative Models (DGMs), such as Diffusion Models (DMs) and Flow Matching (FM) models, have emerged as powerful tools for capturing complex temporal distributions. While DMs offer high-quality samples, they suffer from high computational costs and slow inference due to iterative denoising, often requiring hundreds of network function evaluations (NFEs), especially for long time series [1]. Flow Matching, a newer paradigm, directly optimizes continuous trajectories, offering faster inference and improved scalability. Recent work, such as FM-TS [2] and TSFlow [3], shows its efficiency in both conditional and unconditional settings for time series generation. Still, two major challenges remain: (1) generating high-quality long sequences while maintaining temporal coherence, which is computationally demanding; and (2) improving generative performance through exogenous data integration. Real-world time series are often influenced by external dynamic or static factors. Effectively incorporating such context can greatly enhance realism and utility. This project addresses these challenges by advancing Flow Matching for long time series generation and developing novel strategies to integrate exogenous data into generative models. The efficiency of FM and its continuous formulation make it a compelling direction for scalable and realistic time series modeling. We will evaluate our approach on widely used benchmarks such as ETT [4], Weather [4], and Traffic [5]—datasets that provide diverse temporal dynamics and exogenous variables.

References

[1] Kollovieh, Marcel, et al. "Predict, refine, synthesize: Self-guiding diffusion models for probabilistic time series forecasting." Advances in Neural Information Processing Systems 36 (2023): 28341-28364.

[2] Hu, Yang, et al. "FM-TS: Flow Matching for Time Series Generation." arXiv preprint arXiv:2411.07506 (2024).

[3] Kollovieh, Marcel, et al. Flow Matching with Gaussian Process Priors for Probabilistic Time Series Forecasting. Proceedings of the Thirteenth International

Conference on Learning Representations (ICLR), 2025. https://openreview.net/forum?id=uxVBbSlKQ4.

[4] Zhou, Haoyi, et al. "Informer: Beyond efficient transformer for long sequence time-series forecasting." Proceedings of the AAAI conference on artificial intelligence. Vol. 35. No. 12. 2021.

[5] Wu, Haixu, et al. "Timesnet: Temporal 2d-variation modeling for general time series analysis." arXiv preprint arXiv:2210.02186 (2022).

Details

Student

SB

Sander Bergman

Supervisor

Vlado Menkovski

Secondary supervisor

Mahdi Mehmanchi

Interested?

Get in contact