This repository contains example implementations for centralized models for the Pandemic Response and Forecasting use case for the US/UK PETS Prize Challenges. It contains code to generate features from the synthetic data provided for competitors, as well as implementations of example baselines. It additionally contains sample from the US network and disease states to demonstrate data formatting expectations.
Note that the example data is just meant to demonstrate data formatting expectations for these example implementations. Model performance for models trained just on this data will be quite poor. Replication of the following performance metrics requires running these example implementations on significantly larger portions of the data.
The following AUPRC scores use the development datasets available from the PETs Prize Challenges.
| Dataset | Logistic Regression | Graph Neural Network |
|---|---|---|
| Virginia | 0.045* | 0.008 |
| UK | 0.098** | 0.053** |
* This metric is the result for Virginia when running the model using 50% of the population
** These metrics are the results for UK when running the models using 5% of the population
The requirements for these implementations are listed in requirements.txt
Note that additional steps may be necessary if you wish to train the graph
neural network using GPUs.
Examples of data files used in this implementation may be found in the
sample_data directory. This contains a smaller contact file, contacts.csv,
representing a subset of the edges from the VA graph. This object is a csv
containing a list of edges.
The other files in the sample_data directory are csv files pop.csv,
training.csv and target.csv. These files are subsets of the population,
training and evaluation files containing only the persons present in the
graph.pkl subgraph.
The directory random_walks contains example data produced by the DeepInf/rwr_all.sh
bash script. These are random walks on graph.pkl produced by rwr_all.py.
The directories sample_data/gnn and sample_data/logistic_regression
The logistic regression data generation, training, and evaluation can be executed
by going into the logistic_regression directory and executing the
./logistic_regression.sh bash script.
This generates the training and evaluation data for the logistic regression in
the sample_data/logistic_regression/ directory, and performs the training and
evaluation of the data.
The GNN data generation can be done by going into the DeepInf directory, and
executing the rwr_all.sh bash script followed by the make_data.sh script.
This generates the training and evaluation data for the Graph Neural Network
based example implementation in sample_data/gnn/train/
and sample_data/gnn/eval
To train and evaluate the model, executing the train_gnn.sh bash script.
This creates checkpoint models in the sample_data/gnn/models and a csv with
predictions sample_data/gnn/test_eval.csv
The code for the graph neural network architecture is used from
Jiezhong Qiu et al., “DeepInf: Social Influence Prediction with Deep Learning,” in Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, 2018, 2110–19, https://doi.org/10.1145/3219819.3220077.
To run these models on larger datasets, it may be desirable to parallelize certain aspects of the data preparation. Much of the data preparation is embarrasingly parallel. While the specifics of how this is parallelized depends on the platform, we have written the data preparation scripts to suggest one possible approach.
The scripts DeepInf/rwr_all.sh, DeepInf/make_data.sh', and logistic_regression/logistic_regression.sh, and all contain for-loops over N_PART` iterations. The contents of these for-loops may be executed in parallel,
and the number of jobs may be increased arbitrarily.