CALF: Aligning LLMs for Time Series Forecasting via Cross-modal Fine-Tuning

[中文解读1][中文解读2]

Introduction

CALF (Orignal name: LLaTA) is a novel cross-modal fine-tuing framework that effectively bridges the distribution discrepancy between temporal data and the textual nature of LLMs, as shown in Figure 1.

Figure 1: The t-SNE visualization of pre-trained word token embeddings of LLM with the hidden features from the penultimate layer from GPT4TS, TimeLLM, TEST, and ours of ETTh2 dataset. Current LLM-based methods either use linear layers to project time series to the LLM's feature dimension or employ cross-attention and contrastive learning techniques, which address only the input side and overlook alignment in the deeper layers. Our CALF achieves better alignment through multi-level cross-modal fine-tuning.

To bridge the modality gap between textual and temporal data, we introduce three meticulously designed cross-modal fine-tuning techniques (see Figure 2):

• Cross-Modal Match Module integrates time series and textual inputs through principal word embedding extraction and a cross-attention mechanism, ensuring efficient alignment of the marginal input distribution between time series and text.
• Feature Regularization Loss aligns the outputs of each intermediate layer, ensuring that gradients at every layer are more effectively guided for better weight updates.
• Output Consistency Loss ensures that the output representations of textual and temporal series modalities correspond effectively, resolving discrepancies in the representation space and maintaining consistent semantic context for time series data.

Figure 2: Conceptual illustration of cross-modal fine-tuning technique.

Prerequisites

Before proceeding, ensure Python 3.9 is installed. Install the required dependencies with the following command:

pip install -r requirements.txt

Dataset Preparation

Long-term Forecasting

Acquire datasets from Autoformer. Organize them in the ./datasets directory as shown below:

datasets
├── electricity
│   └── electricity.csv
├── ETT-small
│   ├── ETTh1.csv
│   ├── ETTh2.csv
│   ├── ETTm1.csv
│   └── ETTm2.csv
├── traffic
│   └── traffic.csv
└── weather
    └── weather.csv

Short-term Forecasting

For short-term forecasting, download the M4 datasets from Time-Series-Library. Place the m4 folder within ./datasets.

Preparing Word Token Embeddings

Execute the command below to extract principal components from the word token embeddings:

python pca.py

These embeddings will be saved in ./wte_pca_500.pt.

Model Training

Training scripts are located in the ./scripts folder. For instance, to train the CALF model on the ETTh2 dataset for long-term forecasting, execute:

sh scripts/long_term_forecasting/ETTh2.sh

For short-term forecasting, use:

sh scripts/short_term_forecasting/m4.sh

Post-Training:

• Trained models will be saved in ./checkpoints.
• Numerical results are available in .npy format under ./results.
• Detailed summaries of performance metrics can be found in ./results_{task_name}.txt.

Citation

If this repository contributes to your research, please consider citing our work:

@article{liu2024taming,
      title={CALF: Aligning LLMs for Time Series Forecasting via Cross-modal Fine-Tuning}, 
      author={Liu, Peiyuan and Guo, Hang and Dai, Tao and Li, Naiqi and Bao, Jigang and Ren, Xudong and Jiang, Yong and Xia, Shu-Tao},
      journal={arXiv preprint arXiv:2403.07300},
      year={2024},
      arxiv={2403.07300}
}

Acknowledgements

Our gratitude extends to the authors of the following repositories for their foundational model implementations:

• Autoformer
• Time-Series-Library
• PatchTST
• Crossformer
• One-Fits-All

Contact Us

For inquiries or further assistance, contact us at lpy23@mails.tsinghua.edu.cn or open an issue on this repository.