Using the predictions of the agent-based simulation code epydemy, we train a deep neural-network to help identifying the individual susceptibile to catching the virus (the high-risk group). This predictions for each individual is based on the following features:
- the number of home-mates for that individual,
- the number of co-workers for that individual, and
- the total number of social-place for that individual.
Using these features, we can find the group which are most at danger, e.g. the individual who have 75% chance of being infected (or more).
The deep neural-network is trained by using almost 4 millions individual in 400 simulated cities, for whom the susceptibility to the virus is turned into a probability. With this data, the trained neural-network is capable of identifying the high-risk group by
- precision = 0.90%
- recall = 0.85%
- F1 = 0.88%
The neural-network is build and trained by TensorFlow 2.0.1 .
List of important files:
preprocessing_data.py: converts the raw data of each individual to a probability.trainer.py: builds the neural-network and trains it.
The next step is to use the predictions of this neural-network and selectively quarantine some agents in epydemy to prevent the pandemic from an outbreak.