Comparing supervised learning algorithms and artificial neural networks for conflict prediction: performance and applicability of deep learning in the field

Machine learning algorithms and artificial neural networks promise a new and powerful approach for making better and more transferable predictions in global conflict research. In this paper, a novel conflict dataset for the prediction of conflict intensity is introduced. It includes seven socio-economic and political indicators spanning a set of 851 country-years. This set of indicators is combined with conflict intensity data covering the timeframe of 2009–2015 to build a viable predictor framework. With this dataset as a foundation, a wide range of different predictive methods are tested, including linear discriminant analysis, classification and regression trees, k-nearest neighbor, random forest and several series of advanced artificial neural networks including a novel non-sequential long-short-term memory setup. Acknowledging the potential of deep learning techniques for many disciplines and projects, this paper shows, that for this type of assembled medium sized data, resembling many common research frameworks in Social and Political Sciences, using neural networks as singular approach might not be fruitful. The advantages of neural networks do not always outweigh their practical and technical disadvantages in small or medium data settings. The argument derived from this study is that researchers should combine Supervised Learning Algorithms and Deep Learning Networks as a general approach in similar predictive setups, or carefully evaluate for each dataset and project if the added complexity accompanied with using networks is indeed translating into better predictive performance.