CDL folder aims to run the estimator from Causal Dynamics Learning paper.

In this repository you may find the method explained in [1] to estimate conditional mutual information. This technique is based on variational lower bounds for relative entropy known as Donsker-Varadhan bound (DV bound) and NWJ bound. Below is the DV bound for conditional mutual information:

We use the k- nearest neighbor technique to help us design a neural classifier that is the basis of our estimation.

The model that we used in our simulations is a Gaussian model:

The MI-Diff directory contains my independent implementation of the method proposed in the [2], where the authors estimate conditional mutual information as the difference of two mutual information terms. Among several methods in [2], the MI-Diff method performs better according to the provided simulation results. So we compare our results with this approach.

The neural network is implemented with PyTorch. For k-nearest neighbor, we use Scikit-learn library.

To run the code and reproduce the results in the paper use the help below:

--ker = 1: also run kernel based model (>10mins) --rl = 1: use DGP defined in structurel.py where the true CMI need to be re-calculate (TODO)

python main.py --d 5 --k 20 --n 80000 --scenario 0 --seed 123

python main.py --d 3 --k 10 --n 80000 --scenario 1 --seed 123

python main.py --d 5 --k 10 --n 80000 --scenario 2 --seed 123

python MIDiff.py --d 5 --n 80000 --scenario 0 --seed 123

python MIDiff.py --d 3 --n 80000 --scenario 1 --seed 123

The provided notebook shows how to load and visualize the data

[1] Sina Molavipour, Germán Bassi, Mikael Skoglund, 'On Neural Estimators for Conditional Mutual Information Using Nearest Neighbors Sampling,' arXiv preprint arXiv:2006.07225, 2020.

[2] Sudipto Mukherjee, Himanshu Asnani, Sreeram Kannan, 'CCMI : Classifier based Conditional Mutual Information Estimation,' Proceedings of The 35th Uncertainty in Artificial Intelligence Conference, PMLR 115:1083-1093, 2020.