Multi-scale hypergraph meets LLMs: Aligning large language models for time series analysis
Zongjiang Shang, Dongliang Cui, Binqing Wu, Ling Chen
TL;DR
This paper introduces MSH-LLM, a novel multi-scale hypergraph approach that enhances large language models for time series analysis by capturing multi-scale structures and aligning modalities across different scales.
Contribution
The paper proposes a multi-scale hypergraph framework with hyperedging, cross-modality alignment, and a mixture of prompts to improve LLMs' understanding of time series data.
Findings
Achieves state-of-the-art results on 27 real-world datasets
Effectively captures multi-scale temporal patterns
Enhances LLMs' ability to analyze time series data
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…
Peer Reviews
Decision·ICLR 2026 Poster
1. The paper insightfully identifies the multi-scale structural misalignment between natural language and time as a fundamental bottleneck for applying LLMs to time-series analysis. 2. By introducing learnable hyperedges and a sparsification strategy, the proposed method not only mitigates the noise sensitivity of traditional patch-based approaches but also captures implicit group-level interactions within time series data.
1. The paper does not justify the node–hyperedge similarity in the hyperedging mechanism (Eqs. 2–3), nor does it analyze the theoretical relation between the number of hyperedges M^s and the time-series dimensionality D. 2. The work does not report inference latency on ETTh1 with H=720, so it is unclear whether the method is suitable for real-time scenarios such as power-load dispatch. Memory consumption is also missing; GPU usage on high-dimensional data is not reported, limiting an assessment
1. This paper introduces a hyperedging mechanism to extract group-wise information from multi-scale temporal features, enhancing multi-scale semantic information of time series semantic space as a novel contribution. 2. This paper designs a cross-modality alignment module to perform multi-scale alignment and obtain richer representations, while the mixture of prompts mechanism enhances the reasoning ability of LLMs. 3. This paper conducts comprehensive experiments on 27 real-world datasets acr
1. In Sec 4.1, the word token embeddings U are transformed into U1 through linear mapping. Is the linear mapping learnable or a semantic-distance-based mapping? Will each text prototype possesses explicit meaning? 2. As the pre-trained LLM is freezed during training according to Figure 1, I am confused whether all the proposed modules will be trained simultaneously. If so, what are the training objectives? 3. Experimental results in Table 1 and 2 show minor improvements between MSH-LLM and t
- The paper introduces a multi-scale hypergraph framework to align LLMs for time series analysis. The proposed hyperedging mechanism and mixture of prompts strategy demonstrate better performance beyond existing single-scale or prompt-based methods. - This paper is well-written and easy to follow. - The framework is technically solid and systematically evaluated on 27 real-world datasets covering forecasting, classification, few-shot, and zero-shot learning.
- The provided code files could not be opened successfully, which raises concerns about the reliability of the reported experimental results. - The overall framework appears to combine elements from Time-LLM [1] and hypergraph-based multi-scale modeling. As a result, the methodological novelty may be limited unless the authors can clearly articulate the unique contributions beyond this integration. - The Cross-Modality Alignment (CMA) module used for aligning text and time-series modalities ha
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsTime Series Analysis and Forecasting · Machine Learning in Healthcare · Topic Modeling