👏 The paper is accpeted by CIKM 2025!

🙋 Please let us know if you find out a mistake or have any suggestions!

🌟 If you find this resource helpful, please consider to star this repository and cite our research:

@article{wang2024tabletime,
  title={Tabletime: Reformulating time series classification as zero-shot table understanding via large language models},
  author={Wang, Jiahao and Cheng, Mingyue and Mao, Qingyang and Liu, Qi and Xu, Feiyang and Li, Xin and Chen, Enhong},
  journal={arXiv e-prints},
  pages={arXiv--2411},
  year={2024}
}

In today’s data-driven world, multivariate time series (MTS) are essential in areas like healthcare, industrial monitoring, and behavior recognition. Traditional time series classification (TSC) methods struggle to capture temporal dependencies and multi-channel patterns, while deep learning models, despite their strong performance, are often complex and opaque.

Large language models (LLMs) offer powerful reasoning and cross-domain generalization, but applying them directly to TSC is hindered by data–text mismatches, high costs, and underutilized reasoning ability. To address these issues, we propose TableTime, a framework that reformulates MTS classification as a table comprehension task, enabling LLMs to better exploit their reasoning power and offering a new paradigm for time series analysis.

Whether using traditional classification or deep learning, it's difficult to accurately model both temporal and channel information. In LLM-based methods, aligning the numerical modality with the LLM's textual modality is a crucial consideration.

To model the temporal dependencies in time series as losslessly as possible while preserving channel information and aligning the time series with the LLM's semantic space, the team used a table to model the time series, with the horizontal axis representing time information and the vertical axis representing channel information.

To help LLM better handle unseen samples, the team proposed a Neighbor-Assisted Contextual Reasoning mechanism. This mechanism retrieves neighboring samples from the training data, providing important contextual guidance to the model.

Retrieval methods can include DTW distance, Euclidean distance, Manhattan distance, and others. By retrieving neighbors, LLM can obtain the neighbors and labels of test samples in the training set, thereby enhancing LLM's reasoning capabilities.

The LLM's responses are highly random. To mitigate this, the team proposed Multi-Path Ensemble Enhanced Reasoning. This utilizes multiple different reasoning paths to generate diverse results, improving the model's robustness and accuracy.

To guide the LLM in performing step-by-step analysis, the team implemented a Complex Problem Decomposition Mechanism. This breaks down a complex problem into several smaller ones, guiding the LLM in step-by-step classification, resulting in more accurate analysis.

1, Can Slow-thinking LLMs Reason Over Time? Empirical Studies in Time Series Forecasting

Authors: Cheng, Mingyue and Wang, Jiahao and Wang, Daoyu and Tao, Xiaoyu and Liu, Qi

@article{wang2025can,
  title={Can slow-thinking llms reason over time? empirical studies in time series forecasting},
  author={Wang, Jiahao and Cheng, Mingyue and Liu, Qi},
  journal={arXiv preprint arXiv:2505.24511},
  year={2025}
}

2, FormerTime: Hierarchical Multi-Scale Representations for Multivariate Time Series Classification.

Authors: Cheng, Mingyue and Liu, Qi and Liu, Zhiding and Li, Zhi and Luo, Yucong and Chen, Enhong

@inproceedings{cheng2023formertime,
  title={Formertime: Hierarchical multi-scale representations for multivariate time series classification},
  author={Cheng, Mingyue and Liu, Qi and Liu, Zhiding and Li, Zhi and Luo, Yucong and Chen, Enhong},
  booktitle={Proceedings of the ACM Web Conference 2023},
  pages={1437--1445},
  year={2023}
}

3, InstructTime: Advancing Time Series Classification with Multimodal Language Modeling.

Authors: Cheng, Mingyue and Chen, Yiheng and Liu, Qi and Liu, Zhiding and Luo, Yucong

@article{cheng2024advancing,
  title={Advancing Time Series Classification with Multimodal Language Modeling},
  author={Cheng, Mingyue and Chen, Yiheng and Liu, Qi and Liu, Zhiding and Luo, Yucong},
  journal={arXiv preprint arXiv:2403.12371},
  year={2024}
}

4, TimeMAE: Self-supervised Representation of Time Series with Decoupled Masked Autoencoders.

Authors: Cheng, Mingyue and Liu, Qi and Liu, Zhiding and Zhang, Hao and Zhang, Rujiao and Chen, Enhong

@article{cheng2023timemae,
  title={Timemae: Self-supervised representations of time series with decoupled masked autoencoders},
  author={Cheng, Mingyue and Liu, Qi and Liu, Zhiding and Zhang, Hao and Zhang, Rujiao and Chen, Enhong},
  journal={arXiv preprint arXiv:2303.00320},
  year={2023}
}

5, CrossTimeNet: Learning Transferable Time Series Classifier with Cross-Domain Pre-training from Language Model.

Authors: Cheng, Mingyue and Tao, Xiaoyu and Liu, Qi and Zhang, Hao and Chen, Yiheng and Lei, Chenyi

@article{cheng2024learning,
  title={Learning Transferable Time Series Classifier with Cross-Domain Pre-training from Language Model},
  author={Cheng, Mingyue and Tao, Xiaoyu and Liu, Qi and Zhang, Hao and Chen, Yiheng and Lei, Chenyi},
  journal={arXiv preprint arXiv:2403.12372},
  year={2024}
}