TPP-LLM: Modeling Temporal Point Processes by Efficiently Fine-Tuning Large Language Models

This repository provides the implementation of TPP-LLM, a framework that integrates Temporal Point Processes (TPPs) with Large Language Models (LLMs) for event sequence prediction. The repository includes scripts for training and evaluating the TPP-LLM model on various real-world datasets. For more details on the methodology and experiments, please refer to our paper.

Features

• Semantic and Temporal Modeling: Combines LLMs with TPPs to capture both the semantic richness of event descriptions and the temporal dynamics of event sequences.
• Parameter-Efficient Fine-Tuning: Utilizes Low-Rank Adaptation (LoRA) to efficiently fine-tune the LLM for temporal modeling, reducing computational costs while maintaining high performance.
• Superior Performance: Achieves competitive performance compared to existing models in sequence modeling and event prediction across multiple real-world datasets.

Installation

1. Clone the repository:

   git clone https://github.com/zefang-liu/TPP-LLM
   cd TPP-LLM

2. Install the required dependencies:

   pip install -r requirements.txt

3. Add the source code to your Python path:

   export PYTHONPATH=$PYTHONPATH:<path-to-your-folder>/src

Usage

To train and evaluate the model, use the provided configuration files. For example, to run the model on the Stack Overflow dataset:

python scripts/train_tpp_llm.py @configs/tpp_llm_so.config

You can adjust the configuration to run experiments on other datasets by selecting the appropriate config files located in the configs/ directory.

Datasets

The cleaned data used in this project can be downloaded from Hugging Face. Or you can download and preprocess raw data by running the notebook notebooks/tpp_data.ipynb. Supported datasets include:

• Stack Overflow
• Chicago Crime
• NYC Taxi
• U.S. Earthquake
• Amazon Reviews

Processed datasets will be stored in the data/ directory.

Citation

If you find this code useful in your research, please cite our paper:

@article{liu2024tppllmm,
  title={TPP-LLM: Modeling Temporal Point Processes by Efficiently Fine-Tuning Large Language Models},
  author={Liu, Zefang and Quan, Yinzhu},
  journal={arXiv preprint arXiv:2410.02062},
  year={2024}
}

Questions or Issues

If you have any questions or encounter any issues, please feel free to submit an issue on our GitHub repository.

Acknowledgment

We would like to thank the developers of EasyTPP for their valuable implementation of TPPs, which served as a reference for this work.

License

This project is licensed under the Apache-2.0 License.