baosws / DINE

Diffeomorphic Information Neural Estimation (DINE) - AAAI 2023

Diffeomorphic Information Neural Estimation (DINE)

This is the implementation of our paper: Bao Duong and Thin Nguyen. Diffeomorphic Information Neural Estimation. In AAAI Conference on Artificial Intelligence (AAAI 2023).

Dependencies

conda create -n dine python=3.8
conda activate dine
conda install pytorch -c pytorch
conda install numpy pandas seaborn matplotlib scikit-learn scipy
pip install pytorch-lightning==1.5.3 causal-learn==0.1.3.0 CCIT==0.4 tf-slim tensorflow

Note: causal-learn is for KCIT, CCIT is for CCIT, tensorflow is for CCMI & MIND, tensorflow & tf-slim are for CCMI. Please comment out or remove the lines related to these methods from src/methods/__init__.py if you don't want to install these dependencies.

Demo

from src.data.data_gen import simulate_mi, simulate_cmi, simulate_cit
from src.methods import DINE, DINE_CIT

N = 1000 # sample size
d = 5    # dimensionality of X and Y
dz = 5   # dimensionality of Z

print('-' * 30)
print('Mutual Information (MI) estimation')
X, Y, Z, mi = simulate_mi(N=N, d=d, rho=0.8, random_state=0) # here Z is empty
est = DINE(X=X, Y=Y, Z=Z, random_state=0)
print(f'Ground truth: {mi:.4f}')      # 2.5541
print(f'Estimation:   {est:.4f}\n\n') # 2.5777

print('-' * 30)
print('Conditional Mutual Information (CMI) estimation')
X, Y, Z, cmi = simulate_cmi(N=N, d=d, dz=dz, rho=0.5, random_state=1)
est = DINE(X=X, Y=Y, Z=Z, random_state=0)
print(f'Ground truth: {cmi:.4f}')      # 0.7192
print(f'Estimation:   {est:.4f}\n\n')  # 0.7168

print('-' * 30)
print('Conditional Independence testing')
X, Y, Z, indep = simulate_cit(N=N, d=1, dz=dz, indep=True, random_state=1)
p_value = DINE_CIT(X=X, Y=Y, Z=Z, random_state=0)
print(f'\tConditional Independent: {p_value = :.4f}') # 0.2500

X, Y, Z, indep = simulate_cit(N=N, d=1, dz=dz, indep=False, random_state=1)
p_value = DINE_CIT(X=X, Y=Y, Z=Z, random_state=0)
print(f'\tConditional Dependent:   {p_value = :.4f}') # 0.0000

Running experiments

For example, to run the "Conditional Mutual Information" experiment (Figure 2 in the paper):

python experiments/exp_cmi/main.py --methods DINE KSG CCMI --n_jobs=8

where available methods are DINE, MINE, MIND, KSG, CCMI, KCIT, DINE_CIT and n_jobs is the number of parallel jobs to run.

Modifiable configurations are stored in experiments/exp_*/config/, and result dataframes are stored in experiments/exp_*/results/ after the command is finished.

Citation

If you find our code helpful, please cite us as:

@inproceedings{duong2023diffeomorphic,
	author = {Duong, Bao and Nguyen, Thin},
	booktitle = {Proceedings of the AAAI Conference on Artificial Intelligence},
	pages = {7468-7475},
	title = {Diffeomorphic Information Neural Estimation},
	year = {2023}
}