Graphs are central representations of information in many domains including biological and social networks. Graph visualization is needed for discovering underlying structures or patterns within the data, for example communities in a social network, or interaction patterns between protein complexes. Existing graph visualization methods, however, often fail to visualize such structures, because they focus on local details rather than global structural properties of graphs. We suggest a novel modeling-driven approach to graph visualization: As usually in modeling, choose the (generative) model such that it captures what is important in the data. Then visualize similarity of the graph nodes with a suitable multidimensional scaling method, with similarity given by the model; we use a multidimensional scaling method optimized for a rigorous visual information retrieval task. We show experimentally that the resulting method outperforms existing graph visualization methods in finding and visualizing global structures in graphs.

No abstract