Unsupervised Detection of Persistent Communities in Dynamic Networks with Network Embeddings

Abstract

Many real-world phenomena can be modeled using dynamic networks organized according to an evolving community structure. While traditional community discovery algorithms rely on structural information, recent years have seen a shift in focus towards methods of embedding nodes in dynamic graphs that incorporate temporal information through times- tamps and contextual information through features. This en- deavour is particularly evident in Graph Neural Networks (GNNs). In this paper, we address the importance of node embeddings and features in understanding network dynamics. First, we propose a general framework for extracting dynamic communities from graph embeddings without prior knowl- edge of these communities. We then design a synthetic dy- namic graph generator that produces network data and fea- tures for GNN training. Experiments on synthetic dynamic networks demonstrate that informative features are essential for ensuring the performance of GNNs. By comparing several embedding algorithms on different scenarios, we demonstrate that our method effectively captures the network’s underlying dynamics at both the macro and node levels. Furthermore, we demonstrate that GNN-based embeddings are more effective at capturing global network dynamics. They significantly out- perform conventional embedding or community detection al- gorithms when it comes to detecting changes in dynamics at the node level. Additionally, we tested our method on a real- world dataset of social media interactions, achieving promis- ing results for network visualization.