This project implemented the proposed new architecture Factorized Sum-product Network (FSPN)

Code structure

The overall structure was inspired and adapted from https://github.com/SPFlow/SPFlow. I directly reused some files and functions of this project as well. I thank the SPFlow authors for their excellent code.

This fspn folder contains six sub-folders namely: Algorithm, Data_prepare, Evaluation, Inference, Learning and Structures. Learning contains the key file: structureLearning.py, which learns the FSPN from scratch using the data. Structure defines the set of nodes (sum, product, factorize and leaf).

How to do some simple experiment

Setup the environment using conda (some packages might not support Linux):

conda env create -f environment.yml
conda activate fspn

Navigate to the Evaluation folder and test the training process of FSPN on some toy datasets.

cd fspn/Evaluation
python test_training.py --dataset strong_corr_cat

The test_on_GAS.ipynb file provides a concrete example of how to use FSPN for cardinality estimation of a real-world dataset.

Citation

If you find this project useful please cite our papers:

@article{wu2020fspn,
  title={FSPN: A New Class of Probabilistic Graphical Model},
  author={Wu, Ziniu and Zhu, Rong and Pfadler, Andreas and Han, Yuxing and Li, Jiangneng and Qian, Zhengping and Zeng, Kai and Zhou, Jingren},
  journal={arXiv preprint arXiv:2011.09020},
  year={2020}
}

@article{zhu2020flat,
  title={FLAT: Fast, Lightweight and Accurate Method for Cardinality Estimation},
  author={Zhu, Rong and Wu, Ziniu and Han, Yuxing and Zeng, Kai and Pfadler, Andreas and Qian, Zhengping and Zhou, Jingren and Cui, Bin},
  journal={VLDB},
  year={2021}
}