Neural Blockmodeling for Multilayer Networks

Networks have a rich history of capturing interactions and are widely used in fields ranging from medicine to sociology to computer science. A network where links between nodes may exist on different relations - called a multilayer network - is a flexible and increasingly common type of network since it allows for describing more complex interactions between nodes. Modeling such networks is thus an important task as it allows machines to reason with them intelligently and perform tasks such a link prediction, node classification, and community detection. Two common approaches to modeling multilayer networks are blockmodeling and embedding models. In this paper, we propose a model which fuses these approaches in a neural architecture, allowing it to leverage their respective strengths. To do this, we follow the standard blockmodeling approach of assigning nodes to communities and extending community interactions to node interactions. Community assignment is inspired by embedding models such that each node has an embedding which generates its community assignment. We account for network multilayeredness by mixing a relation embedding with the community interactions vector which acts as a relational modifier on community interactions. We evaluate our model on the three aforementioned tasks using real-world datasets and compare against state-of-the-art models. We find that performance by state-of-the-art models is highly dataset dependent and that our model achieves comparable results. We conclude that blockmodels may be successfully fused with embedding models for network modeling using a neural architecture and that simply mixing a relation embedding with community interactions is a viable way of handling for network multilayeredness.