DualTime: A Dual-Adapter Language Model for Time Series Multimodal Representation Learning

Jiexia YE, Weiqi Zhang, Ziyue Li, Jia Li, Fugee Tsung

Keywords: Multimodal Representation Learning, Time Series, Large Language Model

TL;DR: A novel textual-temporal multimodal learning paradigm and a dual-adapter language model for time series multimodal learning

Abstract:

The recent rapid advancements in language models (LMs) have garnered attention in time series multimodal representation learning. However, existing contrastive learning-based and prompt-based LM approaches tend to be biased, often assigning a primary role to time series modality while treating text modality as secondary.
We classify these approaches under a temporal-primary paradigm, which overlooks the unique and critical task-relevant information provided by the text modality, failing to fully leverage mutual benefits and complementarity of different modalities.
To fill this gap, we propose a novel textual-temporal multimodal learning paradigm that enables either modality to serve as the primary one while being enhanced by the other, thereby effectively capturing modality-specific information and fostering cross-modal interaction. In specific, we design DualTime, a language model composed of dual adapters to implement temporal-primary and textual-primary modeling simultaneously. Within each adapter, lightweight adaptation tokens are injected into the top layers of LM to encourage high-level cross-modal interaction. The shared LM pipeline by dual adapters not only achieves adapter alignment but also reduces computation resources and enables efficient fine-tuning. Empirically, DualTime demonstrates superior performance, achieving notable improvements of 7\% accuracy and 15\% F1 in supervised settings. Furthermore, the few-shot label transfer experiments validate DualTime's expressiveness and transferability.

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: 5460