Résumé : Population behaviours, such as voting and vaccination, depend on social networks. Social networks can differ depending on behaviour type and are typically hidden. However, we do often have large-scale behavioural data, albeit only snapshots taken at one timepoint. We present a method that jointly infers large-scale network structure and a networked model of human behaviour using only snapshot population behavioural data. This exploits the simplicity of a few parameter, geometric socio-demographic network model and a spin based model of behaviour. We illustrate, for the EU Referendum and two London Mayoral elections, how the model offers both prediction and the interpretation of our homophilic inclinations. Beyond offering the extraction of behaviour specific network structure from large-scale behavioural datasets, our approach yields a crude calculus linking inequalities and social preferences to behavioural outcomes. We give examples of potential network sensitive policies: how changes to income inequality, a social temperature and homophilic preferences might have reduced polarisation in a recent election.