🔥 VDSAgent: A PCS-Guided Multi-Agent System for Veridical Data Science Automation

VDSAgents is a multi-agent system guided by the Predictability-Computability-Stability (PCS) principles, designed to automate veridical data science. The system implements a modular and interpretable workflow for data cleaning, feature engineering, modeling, and evaluation. Each phase is managed by an elegant agent, ensuring both functionality and scientific auditability.

🌟 Key Features

• DSLC Data Science Workflow: Automates various stages of data science
• PCS Principles: Ensures system reliability and robustness
• Multi-Dataset Evaluation: Validated across diverse datasets

🙏 Online playground

Special thanks to Silan Hu for providing an online version of the project, which requires no installation. You can access it at https://easy-notebook.silan.tech/.

📦 Installation

Note: This project requires Python 3.9 or above.

Using Conda

conda env create -f environment.yml
conda activate VDSAgents

Using Pip

pip install -r requirements.txt

🚀 QuickStart

To configure the project, you need to fill in the config.py file with your API details:

1. BASE_URL: Set this to the base URL of your API service.
2. API_KEY: Enter your API key here.

These configurations are necessary for the language model to function correctly. Ensure that you have the correct permissions and access rights to use the API.

To use your own dataset, please store your training and test data in the data_science_project directory following this structure:

- data_science_project/
  - <your_project_name>/
    - data/
      - train.csv
      - test.csv

Next, open the run_experiments.py file and modify the following parameters in the main function:

1. COMPETITION_NAME: Set this to your project name, e.g., COMPETITION_NAME = '<your_project_name>'.
2. PROBLEM_DESCRIPTION and CONTEXT_DESCRIPTION: Optionally, update these strings to describe your specific problem and context.

After making these changes, you can run your experiments using the following command:

python run_experiments.py

📈 Framework and Performance

Below are the framework and performance comparison diagrams:

• 
• 

🤝 Contributing

Contributions are welcome! Please feel free to open issues or submit pull requests for any improvements or bug fixes.

📜 License

This project is licensed under the MIT License - see the LICENSE file for details.

👥 Authors

• Yunxuan Jiang: School of Management, Xi’an Jiaotong University; fengjianliu@stu.xjtu.edu.cn
• Silan Hu: National University of Singapore; silan.hu@u.nus.edu. Yunxuan Jiang and Silan Hu have equally contributed to this work.
• Xiaoning Wang: School of Data Science and Media Intelligence, Communication University of China; sdwangxiaoning@cuc.edu.cn
• Yuanyuan Zhang: Beijing Baixingkefu Network Technology Co., Ltd.; zhang.huanzhiyuan@gmail.com
• Xiangyu Chang: School of Management, Xi’an Jiaotong University; xiangyuchang@xjtu.edu.cn

📄 Publication Status

This work is currently under submission for publication. We appreciate your interest and support.