EasyTPP [ICLR 2024]

EasyTPP is an easy-to-use development and application toolkit for Temporal Point Process (TPP), with key features in configurability, compatibility and reproducibility. We hope this project could benefit both researchers and practitioners with the goal of easily customized development and open benchmarking in TPP.

| Features | Model List | Dataset | Quick Start | Benchmark |Documentation |Todo List | Citation |Acknowledgement | Star History |

News

• [02-17-2024] EasyTPP supports HuggingFace dataset API: all datasets have been published in HuggingFace Repo and see tutorial notebook for an example of usage.
• [09-30-2023] We published two textual event sequence datasets GDELT and Amazon-text-review that are used in our paper LAMP, where LLM can be applied for event prediction! See Documentation for more details.
• [09-30-2023] Two of our papers Language Model Can Improve Event Prediction by Few-Shot Abductive Reasoning (LAMP) and Prompt-augmented Temporal Point Process for Streaming Event Sequence (PromptTPP) are accepted by NeurIPS'2023!

Features [Back to Top]

• Configurable and customizable: models are modularized and configurable，with abstract classes to support developing customized TPP models.
• Compatible with both Tensorflow and PyTorch framework: EasyTPP implements two equivalent sets of models, which can be run under Tensorflow (both Tensorflow 1.13.1 and Tensorflow 2.0) and PyTorch 1.7.0+ respectively. While the PyTorch models are more popular among researchers, the compatibility with Tensorflow is important for industrial practitioners.
• Reproducible: all the benchmarks can be easily reproduced.
• Hyper-parameter optimization: a pipeline of optuna-based HPO is provided.

Model List [Back to Top]

We provide reference implementations of various state-of-the-art TPP papers:

Dataset [Back to Top]

We preprocessed one synthetic and five real world datasets from widely-cited works that contain diverse characteristics in terms of their application domains and temporal statistics:

• Synthetic: a univariate Hawkes process simulated by Tick library.
• Retweet (Zhou, 2013): timestamped user retweet events.
• Taxi (Whong, 2014): timestamped taxi pick-up events.
• StackOverflow (Leskovec, 2014): timestamped user badge reward events in StackOverflow.
• Taobao (Xue et al, 2022): timestamped user online shopping behavior events in Taobao platform.
• Amazon (Xue et al, 2022): timestamped user online shopping behavior events in Amazon platform.

Per users' request, we processed two non-anthropogenic datasets

• Earthquake: timestamped earthquake events over the Conterminous U.S from 1996 to 2023, processed from USGS.

• Volcano eruption: timestamped volcano eruption events over the world in recent hundreds of years, processed from The Smithsonian Institution.

  All datasets are preprocess to the Gatech format dataset widely used for TPP researchers, and saved at Google Drive with a public access.

Quick Start [Back to Top]

We provide an end-to-end example for users to run a standard TPP model with EasyTPP.

Step 1. Installation

First of all, we can install the package either by using pip or from the source code on Github.

To install the latest stable version:

pip install easy-tpp

To install the latest on GitHub:

git clone https://github.com/ant-research/EasyTemporalPointProcess.git
cd EasyTemporalPointProcess
python setup.py install

Step 2. Prepare datasets

We need to put the datasets in a local directory before running a model and the datasets should follow a certain format. See OnlineDoc - Datasets for more details.

Suppose we use the taxi dataset in the example.

Step 3. Train the model

Before start training, we need to set up the config file for the pipeline. We provide a preset config file in Example Config. The details of the configuration can be found in OnlineDoc - Training Pipeline.

After the setup of data and config, the directory structure is as follows:

    data
     |______taxi
             |____ train.pkl
             |____ dev.pkl
             |____ test.pkl

    configs
     |______experiment_config.yaml

Then we start the training by simply running the script

import argparse
from easy_tpp.config_factory import Config
from easy_tpp.runner import Runner


def main():
    parser = argparse.ArgumentParser()

    parser.add_argument('--config_dir', type=str, required=False, default='configs/experiment_config.yaml',
                        help='Dir of configuration yaml to train and evaluate the model.')

    parser.add_argument('--experiment_id', type=str, required=False, default='NHP_train',
                        help='Experiment id in the config file.')

    args = parser.parse_args()

    config = Config.build_from_yaml_file(args.config_dir, experiment_id=args.experiment_id)

    model_runner = Runner.build_from_config(config)

    model_runner.run()


if __name__ == '__main__':
    main()

A more detailed example can be found at OnlineDoc - QuickStart.

Documentation [Back to Top]

The classes and methods of EasyTPP have been well documented so that users can generate the documentation by:

cd doc
pip install -r requirements.txt
make html

NOTE:

• The doc/requirements.txt is only for documentation by Sphinx, which can be automatically generated by Github actions .github/workflows/docs.yml. (Trigger by pull request.)

The full documentation is available on the website.

Benchmark [Back to Top]

In the examples folder, we provide a script to benchmark the TPPs, with Taxi dataset as the input.

To run the script, one should download the Taxi data following the above instructions. The config file is readily setup up. Then run

cd examples
python benchmark_script.py

License [Back to Top]

This project is licensed under the Apache License (Version 2.0). This toolkit also contains some code modified from other repos under other open-source licenses. See the NOTICE file for more information.

Todo List [Back to Top]

• New dataset:
  • Earthquake: the source data is available in USGS.
  • Volcano eruption: the source data is available in NCEI.
• New model:
  • Meta Temporal Point Process, ICLR 2023.
  • Model-based RL via TPP, AAAI 2022.

Citation [Back to Top]

If you find EasyTPP useful for your research or development, please cite the following paper:

@inproceedings{xue2024easytpp,
      title={EasyTPP: Towards Open Benchmarking Temporal Point Processes}, 
      author={Siqiao Xue and Xiaoming Shi and Zhixuan Chu and Yan Wang and Hongyan Hao and Fan Zhou and Caigao Jiang and Chen Pan and James Y. Zhang and Qingsong Wen and Jun Zhou and Hongyuan Mei},
      booktitle = {International Conference on Learning Representations (ICLR)},
      year = {2024},
      url ={https://arxiv.org/abs/2307.08097}
}

Acknowledgment [Back to Top]

The project is jointly initiated by Machine Intelligence Group, Alipay and DAMO Academy, Alibaba.

The following repositories are used in EasyTPP, either in close to original form or as an inspiration:

• EasyRec
• EasyNLP
• FuxiCTR
• Neural Hawkes Process
• Neural Hawkes Particle Smoothing
• Attentive Neural Hawkes Process
• Huggingface - transformers

Star History [Back to Top]