High resolution conflict forecasting with spatial convolutions and long short-term memory

Abstract

The 2020 Violence Early Warning System (ViEWS) Prediction Competition challenged participants to produce predictive models of violent political conflict at high spatial and temporal resolutions. This paper presents a convolutional long short-term memory (ConvLSTM) recurrent neural network capable of forecasting the log change in battle-related deaths resulting from state-based armed conflict at the PRIO-GRID cell-month level. The ConvLSTM outperforms the benchmark model provided by the ViEWS team and performs comparably to the best models submitted to the competition. In addition to providing a technical description of the ConvLSTM, I evaluate the model’s out-of-sample performance and interrogate a selection of interesting model forecasts. I find that the model relies heavily on lagged levels of battle-related fatalities to forecast future decreases in violence. The model struggles to forecast escalations in violence and tends to underpredict the magnitude of escalation while overpredicting the spatial spread of escalation.

El concurso de predicciones del sistema de alerta temprana sobre la violencia (Violence Early Warning System, ViEWS) de 2020 desafió a los participantes a producir modelos predictivos de conflictos políticos violentos a altas resoluciones espaciales y temporales. Este documento presenta una red neuronal recurrente de memoria convolucional a corto y largo plazo (convolutional long short-term memory, ConvLSTM) capaz de predecir el cambio de registro en las muertes relacionadas con las batallas como resultado de los conflictos armados de estado a nivel de mes de celda de PRIO-GRID. La ConvLSTM supera el modelo de referencia proporcionado por el equipo de ViEWS y funciona de manera similar a los mejores modelos presentados en el concurso. Además de proporcionar una descripción técnica de la ConvLSTM, analizo el rendimiento del modelo fuera de la muestra y cuestiono una serie de interesantes previsiones del modelo. Considero que el modelo se basa, principalmente, en niveles rezagados de víctimas mortales a causa de las batallas para predecir las futuras disminuciones de la violencia. El modelo se esfuerza por predecir las escaladas de la violencia y tiende a predecir con poca frecuencia la magnitud de la escalada, pero con más frecuencia la propagación espacial de esta.

Le concours 2020 du système d’alerte précoce sur la violence (Violence Early Warning System, ViEWS) a mis les participants au défi de produire des modèles prédictifs des conflits politiques violents à hautes résolutions temporelles et spatiales. Cet article présente un réseau de neurones récurrents à mémoire convolutive à long terme à court terme (ConvLSTM) capable de prévoir l’évolution logarithmique des décès liés aux combats résultant de conflits armés étatiques au niveau Cellule par mois de la grille PRIO. La ConvLSTM surpasse le modèle de référence fourni par l’équipe ViEWS et offre des performances comparables à celles des meilleurs modèles soumis pour le concours. En plus de fournir une description technique de la ConvLSTM, j’évalue les performances hors échantillon du modèle et j’interroge une sélection de prévisions intéressantes du modèle. J’ai constaté que le modèle dépendait fortement des niveaux décalés des décès liés aux combats pour prévoir les futures diminutions de la violence. Le modèle peine à prévoir les escalades de la violence et tend à sous-estimer la magnitude de l’escalade tout en surestimant sa propagation spatiale.

Keywords:

• Forecasting
• machine learning
• neural networks
• political conflict

View correction statement:

Correction

Correction Statement

This article was originally published with errors, which have now been corrected in the online version. Please see Correction (http://dx.doi.org/10.1080/03050629.2022.2101217).

Notes

1 I also make predictions for s = 1 and therefore sometimes refer to seven time steps.

2 In video, the features axis may represent, for example, color channels. In our case, features are PRIO-GRID cell-level covariates.

3 More specifically, the input comprises a tensor with dimensions representing time step, longitude, latitude, and features. The output comprises a tensor with dimensions representing longitude, latitude, and time leads, where the time leads correspond to $s\in \{1,\dots ,7\}.$

4 Here $\text{ln}\hspace{0.17em}(\text{fatalities})$ is shorthand for $\text{ln}\hspace{0.17em}(\mathtt{ged}\_\mathtt{best}\_\mathtt{sb}+1).$

5 Due to an oversight, one cell-level feature included in the benchmark model was omitted from this analysis: spdist_pgd_petroleum, distance to the nearest petroleum resource. Additionally, PRIO provides measures of one-sided and non-state violence similar to the state-based violence measure utilized here. While these are included in the benchmark model, I do not include them in the competition entry model. However, I reestimate the ConvLSTM to include these additional predictors and discuss the results in the section called Test Partition Evaluation.

6 Dropout and batch normalization are often considered forms of regularization, though dropout is also equivalent to data augmentation under certain circumstances (Wager, Wang, and Liang 2013; Zhao et al. 2019). Nevertheless, data augmentation applied to the input images (e.g., skew or rotation) may result in improved predictive performance and could be a promising avenue for future work (Shorten and Khoshgoftaar 2019).

7 Training time is ∼1.5 h.

8 Though this is not shown here and is left for future work.

9 https://github.com/UppsalaConflictDataProgram/OpenViEWS2.

10 $\text{TADDA}=({\sum }_{i=1}^N|{\Delta }_i-\widehat{{\Delta }_i}|+t_d)/N$ where t_d is a penalty, d represents a chosen method for handling near-zero values, and ϵ (not shown) defines “near zero.” ϵ = 1 is chosen following the default in OpenViEWS2.

11 MSE for the model that includes the attention layer is nearly identical to the model without the attention layer.

12 There are some negative predicted values when logged fatalities are 0, but these are very small in magnitude.

13 pgm 127139 did not experience any escalation in violence during the period of test_20–21.

14 This method was developed specifically for this paper but bears some resemblance to the occlusion sensitivity method described by Zeiler and Fergus (2014).

15 Given the noisiness of social processes and data collection, as well as the natural geographic and social heterogeneity of PRIO-GRID cells, I am not convinced that correcting for land area differences of up to 50% will result in substantially improved predictive performance.