unhcr-mauritania/model

Vulnerability modeling # Modeling Assistance Prioritisation This project allows to run an assistance targeting categorisation. Elements from this script were insipired by a Microsoft tutorial The different categories were defined based on the profiling study (**Who?**) and the result of the strategic discussion on targeting (**What?**) and allow to address the question on **How** assistance will be delivered. This specific project focus on situation where 3 prioritisation categories are defined: * "full.allocation", * "reduced.allocation", * "no.allocation" The categories are built from a series of variables collected through a sample-based survey. The objective is to be able to categorise the full population that is registered. The variable to predict comes from the survey while the predictors are the variable collected at registration level. ## Walk-through The information pulled out from the registration database through `get_data_from_db.R`. [^1] Data are then reshapped to prevent imbalances within each modalities of the selected variable. This is done in `feature.R`. The result of a vulnerability survey using measurement of food security is then merged with registration information. The dataset is separated in two parts one for training the model and the second to check its accuraracy, the dataset split allow to minimise the risk of model overfitting In the next step, different models are trained through `multi-classification-modeling.R`. The following algorithms are compiled and compared to generate prediction: * Multinomial Logistic Regression * Neural Network Regression * Gradient Boosted Decision Trees * Random Decision Forest The resulting models can be eventually re-trained based on potential categories revision for cases where the targeting committe would redress cases categorisation. The model is then applied to the registration registry in order to get the categories for each case: `run-prioritisation.R` A script `schedule.R` is available to schedule the cate …