HINet: Heterogeneous Interference Network

This repository provides the code for the paper "Estimating treatment effects in networks by learning exposure mappings".

The structure of the code is as follows:

HINet/
|_ data/
  |_ semi_synthetic/                   
    |_ BC/
    |_ Flickr/
|_ scripts/                  
  |_ main_sweep.py            # Script to run a wandb sweep                   
|_ src/
  |_ data/
    |_ data_generator.py      # Code to generate synthetic and semi-synthetic data
    |_ datatools.py                 
  |_ methods/
    |_ Attention_layer.py
    |_ Causal_models.py       #Implementation of HINet + other methods
    |_ utils.py
  |_ utils/
      |_ metrics.py/          # CNEE and PEHNE implementations
      |_ utils.py/
  |_ training.py              # Trainer class

Installation.

The requirements.txt provides the necessary packages. All code was written for python 3.12.3.

Usage

Download the data for the BC and Flickr datasets from Google Drive. The original Flickr and BC data comes from this repo. We use the same data as Jiang & Sun (2022). Put the data in the data/semi_synthetic/ folder. For the Homophily and BA dataset, set the parameter homophily to True or False, respectively. Now, the results from the paper can be reproduced by selecting the right parameters in the wandb sweep configuration.

Acknowledgements

Our code builds upon the code from Jiang & Sun (2022) and Chen et al. (2024).

Jiang, S. & Sun, Y. (2022). Estimating causal effects on networked observational data via representation learning. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management, (pp. 852–861).

Chen, W., Cai, R., Yang, Z., Qiao, J., Yan, Y., Li, Z. & Hao, Z.. (2024). Doubly Robust Causal Effect Estimation under Networked Interference via Targeted Learning. Proceedings of the 41st International Conference on Machine Learning, in Proceedings of Machine Learning Research 235:6457-6485