University dissertation research project- Causal Implicit GAN, Causal Discovery using a novel GAN structure. For this research, I received a score of 83 (high first). Please note this is an experimental project, so the code is not production-quality.
CIGAN levarages a GAN architecture, trained on ground-truth causal datasets, to generate new causal graphs, as DAGs (directed graphs)
Example of an output from CIGAN
-
It is reccomended that you use a jupyter notebook to get started, you can set up the model like this:
temp_model = CIGAN7.CIGAN(dataset,batch_size,latent_dim,hidden_dim,lr,epochs,sample_interval) -
See the jupyter notebook 'testing.ipynb' for an example on how to run a dataset on the model
-
Some dataset files are included in this repo, but the model must take the data as a numpy array, with no headers, examples on how the data is converted can be found in the 'other' folder
-
Code to visualise the results is availible, and the model can also output images of the results
-
A review of the model on different types of data can be found in an excel file in 'other'
-
Some code and design in based on DAG-WGAN (https://arxiv.org/abs/2204.00387)
-
Please reference my paper and DAG-WGAN if any code is used.