| Abstract |
Generative modelling of inter-group relations enables probabilistic forecasting of possible conflict for scenarios where real-world data is sparse. In order for such models to have relevance and integrity, it is important to ensure that real-world data is used to parameterise the model and verify its characteristics. In this paper we investigate how real-world datasets can be mapped into generative model parameters concerning group structures and behaviours. We highlight the issues involved and present a framework for classifying potential data based on three attributes: (i) inter-group structure, (ii) inter-group actions and (iii) impact of actions. We argue that these attributes are fundamental for benchmarking and developing generative models in the context of limited existing data on inter-group interaction. |