Random Forest Predictions with Dyad Features

Abstract:

We introduce a dyad-centric approach to predict the severity of conflict on the grid level. The main aim of this approach is to address heterogeneity in grid-level predictions that stem from particular armed organization dyads present in a grid and their spatial proximity. Thus, we project dyad specific distributional features to the grid level to address dyad related heterogeneity. Using dynamic-time-warping, we leverage hierarchical clustering to infer different types of severity, both spatial and temporal. Our statistical learning approach to predict the severity of conflict relies on Random Forest approaches for continuous outcomes, known to deal well with non-linearities. We train separate Random Forest models for each t+m month period.

Authors:

Kristian Skrede Gleditsch, Finn L. Klebe, and Nils W. Metternich