This repository provides a reference implementation of CausalMotifs as described in the paper:

Causal Network Motifs: Identifying Heterogeneous Spillover Effects in A/B Tests.
Yuan Yuan, Kristen M. Altenburger, and Farshad Kooti.
TheWebConf'2021.

This repository contains code to replicate the simulation results in the paper. We do not include data for replicating internal FB results but note that the code architecture is the same as the simulation.

Please use Python 3

See requirements.txt for prerequisite Python packages.

• generate_WS.py generates the Watts-Strogatz simulation network discussed in the paper. It outputs two files:

• data_ws.csv: a table file that stores the interference vector, covariates, and outcome.

• probabilities_ws.npy: a numpy file that stores the bootstrapping results. It is a [num_replicates * num_observations * dimension_interference_vector] tensor.

causalPartition.py contains the class of the main algorithm

example.py illustrates the usage. It has the following steps

Data loading and cleaning. It loads the two files generated before, and then clean the data to satisfy positivity requirement

Training. Training the data to generate the partitions (exposure conditions).

• Create a causalPartition class

partition = causalPartition(data_, probabilities_, 'assignment')

• Train and split the space for the interference vector

train_result_nonseparate = partition.split_exposure_hajek(True, outcome, input_features, max_attempt=10, eps=0.01, delta=0.001, criteria={'non_trivial_reduction': 0, 'reasonable_propensity': 0.05})

• Plot the training tree

partition.plot_tree(train_result_separate)

• Sample spliting and plot the estimation tree

est_result_separate = partition.estimate_exposure_hajek(train_result_separate, input_features, outcome, eps=0.01, separate=True)
partition.plot_tree(est_result_separate)

We implemented three different splitting criteria (as the parameter criteria). This is a dict, and please use the key to specify the splitting criterion and the value for the parameter.

• 'non_trivial_reduction': WSSE should be reduced by at least the given parameter ($\gamma$).

• 'separate_reduction': The SSE of a parent node should be greater than both children notes. No paramter needs to be input.

• 'reasonable_propensity': As discussed in the paper footnote (an empirically effective alternative). Needs to input the $\varepsilon$.

• 'min_leaf_size': Leaf size should be greater than the given parameter ($\kappa$)

CausalMotifs is MIT Licensed, as found in the LICENSE file.

• Yuan Yuan, yuan2@mit.edu
• Kristen M. Altenburger, kaltenburger@fb.com
• Farshad Kooti, farshadkt@fb.com

If you find CausalMotifs useful for your research, please consider citing the following paper

@inproceedings{causalmotifs-www2021,
author = {Yuan, Yuan and Altenburger, Kristen M. and Kooti, Farshad},
 title = {Causal Network Motifs: Identifying Heterogeneous Spillover Effects in A/B Tests},
 booktitle = {Proceedings of the 2021 World Wide Web Conference},
 year = {2021}
}

This is one implementation of the method. Possible improvements are also discussed in the paper.