fagan2888 / mixed-migration-forecasting

Forecasting mixed migration for the Danish Refugee Council.

mm4sight

Asset with the Danish Refugee Council for the mixed migration and displacement forecasting project.

Setting up

To get started, clone this repository.

$ git clone [email protected]:IBM/mixed-migration-forecasting.git

Then setup a python environment. If you use the Anaconda distribution (recommended), create a python virtual environment using conda. This way your project specific dependencies are isolated.

$ conda create -n mm4sight python=3.6
$ source activate mm4sight
(mm4sight) $ pip install -r requirements.txt

Here source activate mm4isght activates the virtual environment (source deactivate will deactivate it). pip install directives will install packages for the virtual environment.

Fetch the data artifacts. TBD

Running

The code is in three parts.

Data Transformation (optional): Scripts related to data manipulation. Data transformations are run based on the configuration.json file that has a source file along with a transformer class for each source. All transformer implementations are within the transformer folder. The entire transformation script can be run by activating your python virtual environment and running the wrapper script so.

$ source activate mm4sight
(mm4sight) $ python executor.py

This standardizes the data and populates the processed folder. A base set of processed data is available. So this step is only needed when new data sources or features are added.

Models: The models specification is available as a module in this repository. To recreate the model objects, run

(mm4sight) $ python model/trainer.py

Few additional helper scripts that may be of use.

1. evaluation.py runs the evaluations reported in the paper.
2. gridsearch.py determines grid search parameters for a country level model.
3. test_prediction.py to see how the scoring API can be invoked and how scenarios can be defined.
4. paper-plots.py shows some of the error analysis and plots used in the paper.

Deployment: We use a cloud foundry build pack which can be deployed to IBM Cloud.