Metadata Incorporation

Problem

Different counties have different demographic distributions and geo-spatial properties. Therefore spread of infection varies a lot based on the size of the county, population density, population age, industries, etc. In order to effectively incorporate this into the model we need leverage the following factors:

• Age (mean or median)
• Population density
• Male to Female ratio
• Core industries
• Transportation methods
• Average income
• Racial makeup
• Number of store/shops/bars
• Road networks and city layout.

Data-sources

Generating an embedding

Dummy supervised task

One setup is to train a model on a dummy supervised task then use the resulting intermediate representation. With this method we could train a model to predict the total number of new cases thirty days after the first case. This could for instance give the model an effective representation of how meta attributes relate to the target county. There are many types of these "dummy" supervised tasks we could explore for instance also, however

Unsupervised approaches

** Autoencoders**

Using raw features and fine-tuning

Another option is to simply create an embedding model and train it with the rest of the model. For this we still have a separate model we just train it end-to-end with the rest of the model.

Incorporating embedding

Once we have the emebedding