counterfactual-tpp

This is a repository containing code and real data for the paper Counterfactual Temporal Point Processes, published at NeurIPS 2022.

Pre-requisites

This code depends on the following packages:

1. networkx
2. numpy
3. pandas
4. matplotlib

to generate map plots:

5. GeoPandas
6. geoplot

In order to install the project dependencies you can run the following command:

pip install -r requirements.txt

Code structure

• src/counterfactual_tpp.py: Contains the code to sample rejected events using the superposition property and the algorithm to calculate the counterfactuals.
• src/gumbel.py: Contains the utility functions for the Gumbel-Max SCM.
• src/sampling_utils.py: Contains the code for the Lewis' thinning algorithm (thinning_T function) and some other sampling utilities.
• src/hawkes/hawkes.py: Contains the code for sampling from the hawkes process using the superposition property of tpps. It also includes the algorithm for sampling a counterfactual sequence of events given a sequence of observed events for a Hawkes process.
• src/hawkes/hawkes_example.ipynb: Contains an example of running algorithm 3 (in the paper) for both cases where we have (1) both observed and un-observed events, and (2) the case that we have only the observed events.
• ebola/graph_generation.py: Contains code to build the Ebola network based on the network of connected districts. This code is adopted from the disease-control project.
• ebola/dynamics.py: Contains code for sampling counterfactual sequence of infections given a sequence of observed infections from the SIR porcess (the calculate_counterfactual function). The rest of the code is adopted from the disease-control project, which simulates continuous-time SIR epidemics with exponentially distributed inter-event times.

The directory ebola/data/ebola contains the information about the Ebola network adjanceny matrix and the cleaned ebola outbreak data adopted from the disease-control project.

The directory ebola/map/geojson contains the geographical information of the districts studied in the Ebola outbreak dataset. The geojson files are obtained from Nominatim.

The directory ebola/map/overall_data contains data for generating the geographical maps in the paper, and includs the overall number of infection under applying different interventions.

The directories src/data_hawkes and src/data_inhomogeneous contain observational data used to generate Synthetic plots in the paper. You can use this data to re-generate paper's plots. Otherwise, you can simply generate new random samples by the code.

Experiments

Synthetic

• Inhomogeneous Poisson Processes
• Hawkes Processes

Epidemiological

• Ebola Epidemic Simulation and Counterfactual Calculations
• Generate Geographical Distribution of infections

Citation

If you use parts of the code in this repository for your own research, please consider citing:

@inproceedings{noorbakhsh2022counterfactual,
        title={Counterfactual Temporal Point Processes},
        author={Noorbakhsh, Kimia and Gomez-Rodriguez, Manuel},
        booktitle={Advances in Neural Information Processing Systems},
        year={2022}
}