Simulation studies in DDN3.0 paper

This repository is about the code to perform the simulation and draw the figures related to simulations for DDN 3.0. The main repository of DDN 3.0 is at https://github.com/cbil-vt/DDN3. The results of the simulation are saved at https://zenodo.org/records/10543473.

Installation

Required Python packages: numpy, matplotlib, joblib, rpy2, ddn3, and their dependencies.

Required R packages: JGL, iDINGO, glasso, huge, and their dependencies. We call the huge package to generate synthetic data. JGL and iDINGO contain peer methods.

Users need to Install R locally, and edit the path to R in tools_r.py.

import os
os.environ["R_HOME"] = "C:\\Program Files\\R\\R-4.3.2"

Experiments

Simulations for DDN 3.0 and JGL with different graphs can be performed in exp_step1_batch.py. The script below runs a simulation based on a 200-node graph with two scale-free subgraphs. It also runs methods on a grid of lambda1 and lambda2 values.

import numpy as np
l1_lst = np.arange(0.02, 1.0, 0.02)
l2_lst = np.arange(0, 0.16, 0.025)
graph_type = "scale-free-multi"
batch_run(l1_lst, l2_lst, n1=200, n2=200, n_node=200, ratio_diff=0.25, graph_type=graph_type, n_group=2, n_rep=20)

The simulation with imbalanced samples is in exp_step1_batch_imbalance.py. This script is similar to exp_step1_batch.py. These results can be plotted by exp_step2_analysis.ipynb and the summary plot used in the paper is generated by exp_step2_analysis_summary.ipynb. You can download the simulation results at https://zenodo.org/records/10543473 to draw the figures. Note that you need to change the path in exp_step2_analysis.ipynb and exp_step2_analysis_summary.ipynb to point the downloaded results. The experiments related to DINGO are in exp_step0_ddn_jgl_dingo.ipynb.

The speed comparisons can be found in exp_speed_comparison.ipynb. We found that there is negligible cost for calling R from Python. In addition, the first time we call Numba accelerated code, like DDN 3.0, it will compile the code and it will take some time. This only needs to be done once. Therefore, we record the time for later calls of DDN 3.0.

For more details, leave an issue, or email Yizhi Wang (yzwang@vt.edu).