tram-dag

Code to reproduce the experiments for the TRAM-DAG paper.

Comparison with Existing Normalizing Flows (NFs)

The code to reproduce experiments comparing TRAM-DAG with existing normalizing flows (NFs) is in the folder comparison.

Figures Comparing to VACA

To generate figures for L1 (observational) and L2 (interventional) comparisons on the VACA dataset, use the file comparison/Figure_Triangle_Linear_Bimodal.R.

Figures Comparing to CAREFL

Figure 5, titled "Results of the counterfactual queries as posed in CAREFL," can be reproduced using comparison/carefl_fig5.r.

Simulation Study

The code to reproduce the simulation study is in the folder summerof24. This study explores TRAM-DAG performance on continuous and ordered categorical data types.

Figures and Data for the Continuous Case

Figures and data for the continuous case are generated using triangle_structured_continuous.R. Set specific command-line arguments to define the function form and model type. Alternatively, modify args <- c(1, 'ls') around line 7.

• Example for Figure 7 (Complex TRAM and Complex DGP): set args <- c(4, 'cs').

Figures and Data for Ordered Data Types

Figures for the ordered data type experiments are generated with triangle_structured_mixed.R. To reproduce Figure 8, use args <- c(1, 'ls').

System and Package Information for TRAM-DAG

Key Python Libraries

The experiments were conducted with the following versions:

• TensorFlow: 2.13.0
• TensorFlow Probability: 0.21.0

R Environment

• R version: 4.2.3 (2023-03-15)

Session Information

A detailed record of all R packages, versions, and system settings is saved in session_info.txt for full reproducibility.