BRIDGE: Bootstrapping Text to Guide Time-Series Generation via Multi-Agent Iterative Optimisation and Diffusion Modelling
Keywords: Time Series Generation; AI Agent
TL;DR: We explore the possibility of using text to control time series generation. A multi-agent system is used to optimise text to construct potential datasets iteratively. A ts diffusion model is used for controllable time series generation.
Abstract: Time-series Generation (TSG) is an impactful research direction, as generating realistic sequences can be used to create educational materials, in simulations and for counterfactual analysis in decision making. It has further the potential to alleviate the resource bottleneck that arises from a lack of diverse time-series data required to train large time-series foundational models. However, most existing TSG models are typically designed to generate data from a specified domain, which is due to the large divergence in patterns between different real-world TS domains. In this paper, we argue that text can provide semantic information (including cross-domain background knowledge and instance temporal patterns) to improve the generalisation of TSG. To do so, we introduce ``Text Guided Time Series Generation'' (TG$^2$)---the task of generating realistic time series from handful of example time series paired with their textual description. We further present a Self-Refine-based Multi-Agent LLM framework to synthesise a realistic benchmark for TG$^2$ and show that the collected text descriptions are both realistic and useful for time-series generation. We develop a first strong baseline for the TG$^2$, Bridge, which utilises LLMs and diffusion models to generate time series which encode semantic information as cross-domain condition. Our experimental results demonstrate that Bridge significantly outperforms existing time-series generation baselines on 10 out of 12 datasets, resulting in data distributions that are more closely aligned to target domains. Using the generated data for training positively impacts the performance of time series forecasting models, effectively addressing training data limitations. This work bridges the gap between LLMs and time series analysis, introducing natural language to help the time series generation and its applications.
Primary Area: learning on time series and dynamical systems
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 3688
Loading