Preamble

If you're looking at this repository, there's a good chance you should instead look at the method described here: https://arxiv.org/abs/2205.08033 (This second paper describes how to estimate contagion effects, which is the more obvious case for causal estimation with network structure)

Introduction

This repository contains software and data for "Using Embeddings to Correct for Unobserved Confounding in Networks " (arxiv:1902.04114). The paper describes a method to use network structured data to correct for unobserved confounding in causal inference. For example, a social network carries information about the people in it. We want to assess the effect of some treatment on people in the network using observational data. In general, treatment assignment and outcome may be rely on latent attributes of the people. We show how to use the network itself to correct for such unobserved confounding.

This software builds on relational ERM https://github.com/wooden-spoon/relational-ERM. See that repository for detailed instructions on building predictive models using network embedding methods.

Requirements

1. Python 3.6 with numpy and pandas
2. Tensorflow 1.11
3. gcc

Setup

Run the following command in src to build the graph sampler tensorflow ops:

python setup.py build_ext --inplace

Data

We include pre-processed Pokec data for convenience. Raw data from https://snap.stanford.edu/data/soc-Pokec.html We include the pre-processing scripts in the release; these can be modified as required.

Reproducing the experiments

The default settings for the code match the settings used in the paper. These match the default settings used by relational ERM (i.e., we didn't tune anything).

You'll run the code from src as ./relational_ERM/submit_scripts/run_model.sh Changing flags in this file will replicate different experiments. The simulation setting is controlled by the --simulated flag. Options are attribute ('attribute') or propensity based ('propensity') simulation. The later can be used to reproduce the exogeneity experiments.

The default parameters in the shell script use the region based simulation with Beta=1.0

To reproduce the two-stage training, run with embedding_trainable=false

Finally, the effect estimates can be reproduced by running python -m effect_estimates.compute_single_ate tsv_file_path. This takes in the predictions of the relational erm model (in tsv format) and passes them into downstream estimators of the causal effect.

Misc.

The experiments in the paper initialize from node embeddings that were pre-trained using a purely unsupervised objective. To recreate the initialization embeddings, run run_unsupervised.sh. Then, uncomment --init_checkpoint=$INIT_FILE in run_model.sh