Abstract
Graph representation learning has become a prominent tool for thecharacterization and understanding of the structure of networks in general andsocial networks in particular. Typically, these representation learningapproaches embed the networks into a low-dimensional space in which the role ofeach individual can be characterized in terms of their latent position. A majorcurrent concern in social networks is the emergence of polarization and filterbubbles promoting a mindset of "us-versus-them" that may be defined by extremepositions believed to ultimately lead to political violence and the erosion ofdemocracy. Such polarized networks are typically characterized in terms ofsigned links reflecting likes and dislikes. We propose the latent Signedrelational Latent dIstance Model (SLIM) utilizing for the first time theSkellam distribution as a likelihood function for signed networks and extendthe modeling to the characterization of distinct extreme positions byconstraining the embedding space to polytopes. On four real social signednetworks of polarization, we demonstrate that the model extractslow-dimensional characterizations that well predict friendships and animositywhile providing interpretable visualizations defined by extreme positions whenendowing the model with an embedding space restricted to polytopes.