Neural Autoregressive Flows for Markov Boundary Learning (FANS)

This is the implementation for our paper: Neural Autoregressive Flows for Markov Boundary Learning, accepted at IEEE ICDM 2025.

Setup • Structure • Experiments • Citation

Setup

conda env create -n fans --file env.yml
conda activate fans

Files

• data/: Create synhetic datasets, Masking mechanisms.
• data_gen/: Save data here.
• model/: Source code of models.
• model_save/: Save model checkpoints here.
• result_save/: Save results here.
• utils/: Code utils.
• metrics.py: Metrics for evaluation.
• run_linear.py: Run linear data.
• run_nonlinear.py: Run nonlinear data.

Experiments

Linear data

Default: Data sampled from an ER graph with 100 nodes, expected degree 1, 5000 samples, and Gaussian noise.

python run_linear.py

Nonlinear data

Default: Data are sampled from an ER graph with 30 nodes and expected degree 1, using 1000 samples. The data-generating process is drawn from a Gaussian Process, with additive Gaussian noise.

Train FANS:

python run_nonlinear.py --train --mode flow --d 30 --data_seed 42

Citation

@inproceedings{nguyen2025neural,
      title={Neural Autoregressive Flows for Markov Boundary Learning}, 
      author={Nguyen, Khoa and Duong, Bao and Huynh, Viet and Nguyen, Thin},
      year={2025},
      booktitle = {IEEE International Conference on Data Mining},
}