fetch's Introduction

FETCH

Introduction

Code for "Learning a Data-Driven Policy Network for Pre-Training Automated Feature Engineering "

Accepted in ICLR 2023 Conference

This is an automated feature engineering framework "FETCH", implemented in PyTorch. Data can be accessed through URL in the paper.

How to run it

pip install -r requirements.txt: install needed packages under the environment settings by pip.

python main_attention.py: after specifying the dataset, cuda, and other parameters, you can run FETCH to automate feature engineering for Random Forest or other pre-defined model.

. run_table_1.sh: run Experiment 4.2.

. run_table_3.sh: run Experiment C.3 on large-scale dataset (> 50k rows).

. run_eval_pretrain.sh: run Experiment 4.3 on evaluating pre-trained models.

. run_diff_models.sh: run Experiment 4.4 and C.5 on different models.

. run_table_5.sh: run Experiment C.6.

Environment

The code in this repository is designed and highly recommended to be run on Ubuntu 20.04 or other Linux systems.

Note that running the code with multiple processes (args.worker > 1) on Windows may encounter issues with variable sharing. If you are using Windows, consider setting up a Linux environment (e.g., using a virtual machine or WSL) to run the code.

Citation

@inproceedings{li2023learning,
  title={Learning a Data-Driven Policy Network for Pre-Training Automated Feature Engineering},
  author={Li, Liyao and Wang, Haobo and Zha, Liangyu and Huang, Qingyi and Wu, Sai and Chen, Gang and Zhao, Junbo},
  booktitle={The Eleventh International Conference on Learning Representations},
  year={2023}
}