Field-wise Learning for Multi-field Categorical Data

This repository is the official implementation of Field-wise Learning for Multi-field Categorical Data.

Requirements

The code has been tested with:

• Python 3.6.8
• PyTorch 1.1.0
• lmdb 0.96
• tqdm 4.32.1

Training and Evaluation

1. Download the Avazu and Criteo datasets.

2. To train and evaluate the model(s), run following command (see full input arguments via python run_fwl.py --help):

   python run_fwl.py --dataset-path <path_to_data>
   

For example, to train the model on Criteo datasets, run:

python run_fwl.py  --dataset-path ./data/criteo/train.csv --ebd-dim 1.6 --log-ebd --lr 0.01 --wdcy 1e-6 --include-linear --reg-lr 1e-3 --reg-mean --reg-adagrad

to train the model on Avazu datasets, run:

python run_fwl.py  --dataset-path ./data/avazu/train.csv --ebd-dim 10 --lr 0.05 --wdcy 1e-8 --reg-lr 1e-6 --reg-mean 

Citation

If you find this repository helpful, please consider to cite the following paper:

@inproceedings{NEURIPS2020_70789713,
 author = {Li, Zhibin and Zhang, Jian and Gong, Yongshun and Yao, Yazhou and Wu, Qiang},
 booktitle = {Advances in Neural Information Processing Systems},
 editor = {H. Larochelle and M. Ranzato and R. Hadsell and M. F. Balcan and H. Lin},
 pages = {9890--9899},
 publisher = {Curran Associates, Inc.},
 title = {Field-wise Learning for Multi-field Categorical Data},
 url = {https://proceedings.neurips.cc/paper/2020/file/7078971350bcefbc6ec2779c9b84a9bd-Paper.pdf},
 volume = {33},
 year = {2020}
}