TS-FourierLLM: Frozen Frequency-Domain Large Language Blocks for Enhancing Time-Series Modeling

  • Pengfei Wang
  • , Huanran Zheng
  • , Wenjing Yue
  • , Xiaoling Wang*
  • *Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

Designing effective time-series models is challenging due to factors like complex temporal dependencies, low information density, and scalability constraints. To address these challenges, we propose TS-FourierLLM, a novel approach that integrates a frozen Large Language Models (LLMs) block as a plug-and-play frequency-domain enhancer for time-series modeling. By transferring high-level pre-trained LLM knowledge into the frequency domain, our method bridges the modality gap while avoiding fine-tuning, preserving computational efficiency. The frozen LLM block captures inter-frequency dependencies and enhances global feature representations, complementing time-series encoders. TSFourierLLM achieves up to a 3% performance improvement over state-of-the-art methods on the benchmark. These results demonstrate the effectiveness of utilizing frozen LLMs as modular and task-agnostic components for advancing time-series modeling.

Original languageEnglish
Title of host publicationDatabase Systems for Advanced Applications - 30th International Conference, DASFAA 2025, Proceedings
EditorsFeida Zhu, Ee-Peng Lim, Philip S. Yu, Akiyo Nadamoto, Kyuseok Shim, Wei Ding, Bingxue Zhang
PublisherSpringer Science and Business Media Deutschland GmbH
Pages757-766
Number of pages10
ISBN (Print)9789819538294
DOIs
StatePublished - 2026
Event30th International Conference on Database Systems for Advanced Applications, DASFAA 2025 - Singapore, Singapore
Duration: 26 May 202529 May 2025

Publication series

NameLecture Notes in Computer Science
Volume15987 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference30th International Conference on Database Systems for Advanced Applications, DASFAA 2025
Country/TerritorySingapore
CitySingapore
Period26/05/2529/05/25

Keywords

  • Frequency-Domain Learning
  • Large Language Models (LLMs)
  • Time-Series Modeling
  • Transfer learning

Fingerprint

Dive into the research topics of 'TS-FourierLLM: Frozen Frequency-Domain Large Language Blocks for Enhancing Time-Series Modeling'. Together they form a unique fingerprint.

Cite this