RuDi

Codes and data for RuDi: Explaining Behavior Sequence Models by Automatic Statistics Generation and Rule Distillation. (arXiv)

Usage

Un-tar datasets. (Due to the large size, we do not include the processed Elo. One can download the raw dataset and use our provided pre-processing code.)

tar zxf data.tar.gz

Run the algorithm:

chmod +x run.sh
./run.sh data/vews_all cuda:0 gru

Then the statistics and rules are saved in data/[dataset]/[teacher]_stats and data/[dataset]/[teacher]-rudi_rules.

Dependencies

Our codes work perfectly with the followling packages:

• python=3.8.3
• pandas==1.2.3 (Important: other versions of pandas are likely to raise unexcepted errors)
• numpy==1.21.2
• pytorch==1.6.0
• scikit-learn==0.23.2
• scipy==1.7.3
• tqmd==4.48.2

Datasets

Raw datasets can be downloaded from the following links.

• VEWS: https://cs.stanford.edu/~srijan/vews/vews-raw-dataset.zip
• Elo: https://www.kaggle.com/c/elo-merchant-category-recommendation/
• RedHat: https://www.kaggle.com/c/predicting-red-hat-business-value/

The pre-processing codes are included in data_preprocessing.

Cite us

@inproceedings{zhang2022rudi,
  title={RuDi: Explaining Behavior Sequence Models by Automatic Statistics Generation and Rule Distillation},
  author={Zhang, Yao and Xiong, Yun and Sun, Yiheng and Shan, Caihua and Lu, Tian and Song, Hui and Zhu, Yangyong},
  booktitle={31st ACM International Conference on Information and Knowledge Management},
  year={2022}
}