What drives the VIEWS forecasts?

Using surrogate models to peek into the “black box” of the VIEWS conflict prediction model. 
Understanding how advanced AI and machine learning models make certain predictions is challenging. Typically, models with more complexity perform better than simpler ones, but this complexity makes them harder to interpret. This, in turn, can hinder the extraction of detailed information, which is often relevant for efficient and timely anticipatory action.
To address this issue, various solutions for enhancing the interpretability of complex AI models are being developed. These solutions will eventually be integrated into the VIEWS framework. In the meantime, we have developed a series of “interpretation” models. These models help users asses what input data variables, or broader “themes” thereof, have the most substantial isolated impact on our predictions. We refer to these as our surrogate models.
It is important to note that these models provide only a partial insight and are not designed to indicate causal relationships.

Intro

What are the surrogate models?

The VIEWS surrogate models have been developed to aid interpretation of each of our ensemble model predictions – the monthly VIEWS predictions. They assist interpretation by extracting relevant information from the models, specifically concerned with relationships between in- and output data. They relate selected input features to the final VIEWS model predictions, assessing how much the of main VIEWS conflict prediction can be explained by a selected variable on its own.
  • Please note that interpretations of the surrogate model results are statements about the model, not the real world. They explore the explanatory power of input variables in relation to the VIEWS model prediction, offering information on correlation – not causation.

AVAILABLE DATA

Results from the following surrogate models are currently available in the VIEWS API:
  • Conflict history (surrogate_mean_ch)
  • Infant mortality rate (surrogate_mean_imr)
  • Neighbourhood conflict history (surrogate_mean_nch)
  • Population size (surrogate_mean_pop)
  • Liberal democracy index (surrogate_mean_dem)
The variables above contain data presented as the number of predicted fatalities from each surrogate model. A logged version of the same data is also available for each variable, provided with the variable IDs above plus the suffix _ln. 
Explore the latest data

Method

How do the surrogate models work?

In our implementation, the main VIEWS prediction is used as the dependent variable and one or several predictor(s) as the only independent variable(s) in a linear regression model. If the relationship between the variable and the prediction is not a straight line, we use a “generalized additive model” (GAM) to capture non-linear patterns. This is done for all predictions at the country level. All other predictor variables are ignored, except for the selected one(s).
In a second step, the variance is used to generate separate predictions for each country in the VIEWS model’s scope, expressed as predicted fatalities for a given country and month. These are the results that we make available as our monthly surrogate model predictions in the VIEWS API.
For example, results from the conflict history surrogate model for a given country and month can be interpreted as how much of the overall predicted fatalities (from the main VIEWS model prediction) for the selected month that can be explained solely by the history of fatal conflict in the country at hand.
  • Please note that the result from the different surrogate models do not add up to the total number of fatalities predicted by the main VIEWS model. The models isolate the effect of one or a selected number of features, ignoring all other contributing factors. The results should therefore be seen as a relative indication of the importance of the selected indicators, rather than actionable predictions of their own.
Example of a visualization of the "Variance explained". The plot shows how the (log) number of fatalities predicted six months into the future relate to the (log) number of deaths in state-based conflict up to the last month of data available when the prediction was made (here Jan 2022). Here, conflict history can explain 91.5% of the main VIEWS prediction.
Example of a visualization of the “Variance explained”. The plot shows how the (log) number of fatalities predicted six months into the future relate to the (log) number of deaths in state-based conflict up to the last month of data available when the prediction was made (here Jan 2022). Here, conflict history can explain 91.5% of the main VIEWS prediction.
Example of a surrogate model prediction map for Africa and the Middle East. The plot displays the conflict history surrogate model predictions for March 2024, i.e. the predicted fatalities in March 2024 when looking only at the history of armed conflict up to the last month of input data (here Dec 2023) informing the prediction at hand. The underlying data is available via the VIEWS API.
Example of a surrogate model prediction map for Africa and the Middle East. The plot displays the conflict history surrogate model predictions for March 2024, i.e. the predicted fatalities in March 2024 when looking only at the history of armed conflict up to the last month of input data (here Dec 2023) informing the prediction at hand. The underlying data is available via the VIEWS API.

Learn more in the technical report on the fatalities002 model

Read the report