Graph Visualization with Latent Variable Models
Large graph layout design by choosing locations for the vertices on the plane, such that the drawn set of edges is understandable, is a tough problem. The goal is ill-defined and usually both optimization and evaluation criteria are only very indirectly related to the goal. We suggest a new and surprisingly effective visualization principle: Position nodes such that nearby nodes have similar link distributions. Since their edges are similar by definition, the edges will become visually bundled and do not interfere. For the definition of similarity we use latent variable models which incorporate the userâs assumption of what is important in the graph, and given the similarity construct the visualization with a suitable nonlinear projection method capable of maximizing the precision of the display. We finally show that the method outperforms alternative graph visualization methods empirically, and that at least in the special case of clustered data the method is able to properly abstract and visualize the links.