The Dominance of Geometric Graph Models in Animal Social Networks

Detecting patterns in animal social behaviour and movement is complicated by the diversity of ecological, evolutionary, environmental, and biological drivers of these behaviours such as migration, foraging, assortative mixing, socio-ecological factors, and human influence. Addressing these complexities requires a multidisciplinary approach. Recent advances in network analysis and machine learning offer powerful tools for examining and interpreting complex network structures, aiding in the identification and quantification of movement patterns and the prediction of behavioural changes. Here we use a comparative approach leveraging network analysis and machine learning techniques to assess commonalities in standard theoretical social structures governing networks across the animal kingdom. We investigate how these theoretical structures explain social organization at different scales, from entire populations to smaller groups. By leveraging interpretable machine learning techniques, we examine the predictive power of species and network construction techniques in predicting structural features of animal social networks. We found that geometric graphs are the frequently predicted network model across the animal kingdom. These graphs represent both spatial and social processes and are formed by positioning individuals uniformly on a 2D plane, with links established based on proximity within a specified distance. In particular, geometric graphs demonstrate structural similarities with interaction types and data collection methods. For example, we found that this graph model had strong structural similarities with networks derived from physical contact and spatial proximity data. Networks with small-world properties, in contrast, were rare across all interaction types and collection methods. Additionally, the occurrence of these networks is influenced by the identity of species and sampling duration. Although incorporating species identity into the classification model did not improve the accuracy of the prediction, it enabled us to account for the varying dependencies of biological characteristics on specific behaviours. This study highlights the value of predictive modelling for uncovering ecological drivers of animal network structures. By focusing on standard theoretical models that are well established in network science, we connect animal social networks to broader theoretical findings, while recognizing that more tailored models combining multiple generative models or network properties may offer deeper insights into network structures. We also emphasize the importance of considering how the specific methods used to build networks for each taxon could influence the biological inferences that can be drawn from those networks.