Abstract

Community detection or node clustering is an important topic in network analysis. Many unsupervised clustering algorithms have been developed over the time, such as the Louvain method or Infomap, which can incorporate additional values (or measurements) associated with the graph. More recently, it has also become popular to approach node embedding using graph neural networks.  

Meanwhile, we propose to view the problem as one of classifying whether any existing edge in the graph is network-defining or spurious, where an edge may be said to be ‘spurious’ in this context if it actually connects two nodes that are not members of the same network. The large number of existing unsupervised node clustering methods are no longer effective from this perspective. 

We investigate a supervised approach to network classification by building predictive models of the target network-edges that can be observed in a sample of graph components. We also discuss the possibility of adapting similar node embedding approaches for statistical modelling of valued subgraph motifs, which can be relevant for advanced methods of analysis or micro simulation of complex graph-structured populations.