Learning semi-Markovian DAGs with flow-based VAE

• Project proposed for M1399_000400-2023fall(Deep Learning) at Seoul National University

Overview

• Learning Non-Gaussian linear SEMs with independent noise using IAF based VAE
• Learning linear SEMs with dependent noise (i.e. semi-Markovian graph) using IAF based VAE

Environment

• Python 3.8
• Pytorch 2.1.0
• networkx 2.8.7

Usage

To train the model, run the following command

• Semi-Markovian DAGs

python train.py --dependence_type=1 --dependence_prop=0.3 --node_size=20 --seed=123 --flow_type='IAF'

• Non-Gaussian DAGs

python train.py --dependence_type=0 --graph_dist='laplace' --node_size=20 --seed=123 --flow_type='IAF'

Possible noise_dist : 'normal', 'uniform', 'exponential', 'laplace', 'gumbel'

Paper

• paper Link

Code Sources

Code benefit from the following works

• DAG NOTEARS
• DAG-GNN
• Flow-VAE
• IAF-VAE
• VAE-VPFLOWS