Visualizing the Temporal Similarity Between Clusters of Dynamic Graphs

The evolution of graph structures in large time-varying graphs is often difficult to visualize and interpret due to excessive clutter from overlapping nodes and edges. With limited display area, visual clutter often increases and makes it difficult to recognize developing patterns in embedded subgraphs. In such situations viewers are often hampered in observing and exploring significant changes of the graph components. This poses a cognitive barrier in the visual analytics of large dynamic structures. Another important problem in visualizing dynamic graphs is capturing the difference between graph states. Their state changes often become intractable. In this paper we propose to construct cognitive templates for grouping closely related entities using community detection techniques. The induced subgraphs are collapsed into meta-nodes in order to simplify the representation of large graphs and induce similarities between communities. In order to compute the new structures, we introduce the GCN, or Graph Convolution Network, that learns the representations of sub-graphs induced by communities. The pair-wise similarities can then be calculated by graph-based cluster search algorithms. Furthermore, the proximity state might change temporally. We need to extract the matched communities between consecutive snapshots. Using multi-dimensional scaling and color mappings, we reveal the evolution of graphs at the community level. We evaluate the effectiveness of our method by applying it to the Wikipedia edit history data set.