Multi-scale hypergraph meets LLMs: Aligning large language models for time series analysis

TL;DR

MSH-LLM tackles the challenge of aligning frozen large language models with time series by introducing a triad of innovations: a multi-scale hypergraph-based hyperedging mechanism, a cross-modality alignment (CMA) module, and a mixture of prompts (MoP) to enrich context and reasoning. The framework maps time-series data and language into multi-scale representations, then fuses them through scale-wise hyperedges and cross-attention, guided by learnable prompts and data/task descriptions. Empirical results across 27 real-world datasets in long-term/short-term forecasting, classification, few-shot, and zero-shot learning establish consistent state-of-the-art performance and demonstrate robustness to domain shifts and data scarcity. The work suggests that explicit multi-scale cross-modal alignment and prompt engineering can effectively unlock LLM capabilities for time-series analysis, with practical implications for cross-domain forecasting and decision support.

Abstract

Recently, there has been great success in leveraging pre-trained large language models (LLMs) for time series analysis. The core idea lies in effectively aligning the modality between natural language and time series. However, the multi-scale structures of natural language and time series have not been fully considered, resulting in insufficient utilization of LLMs capabilities. To this end, we propose MSH-LLM, a Multi-Scale Hypergraph method that aligns Large Language Models for time series analysis. Specifically, a hyperedging mechanism is designed to enhance the multi-scale semantic information of time series semantic space. Then, a cross-modality alignment (CMA) module is introduced to align the modality between natural language and time series at different scales. In addition, a mixture of prompts (MoP) mechanism is introduced to provide contextual information and enhance the ability of LLMs to understand the multi-scale temporal patterns of time series. Experimental results on 27 real-world datasets across 5 different applications demonstrate that MSH-LLM achieves the state-of-the-art results.

Figures (8)

• Figure 1: The framework of MSH-LLM. (a) and (b) provide detailed delineation of the multi-scale extraction module, while (c) elaborates on the hyperedging mechanism.
• Figure 2: Prompt example. < > and < > are task-specific configurations and input statistic information, respectively.
• Figure 3: Time series classification results. The results are averaged from 10 subsets of UEA. Higher values mean better performance. Full results are given in Appendix \ref{['TSC']}.
• Figure 4: The impact of different hyperparameters.
• Figure 5: The t-SNE visualization of the output generated by pre-trained LLMs under different prompts.