Points of view: Networks

systems (for example, imagining connections as forces or springs) will often produce visible aggregates of nodes, making it easier to spot hubs and clusters (Fig. 1c). Node-link diagrams can be highly useful but unfortunately do not scale well. As a dataset becomes larger, the visual complexity that results from the added information density approaches an incomprehensible ‘hairball’. For larger undirected networks, ‘adjacency matrices’ are a practical solution (Fig. 2). In this compressed representation, every node in the network is shown as a row and a column with the order of nodes being the same on both axes. A link between two nodes is indicated by filling the two corresponding cells at the intersections of the nodes (Fig. 2a). In this way, adjacency matrices do not suffer from the data occlusions and edge crossings synonymous with nodelink diagrams. One drawback, however, is that adjacency matrices make it difficult to understand the relationships between two nodes that are not directly connected. To maximize the utility of adjacency matrix visualizations, reorder the nodes such that as many filled cells appear next to each other as possible. The result is that clusters are evident as marks near the diagonal and connections ‘between’ clusters appear as clumps away from the diagonal. Similarly, hubs are seen as rows and columns with many filled cells (Fig. 2b). There may be times when both node-link diagrams and adjacency matrices are inadequate for the size of the network. In these in stances, it may be useful to limit the representation to a partial network or rely on relevant statistical measures. For example, a clustering coefficient can be computed that describes the extent of interconnectivity in the neighborhood of a node. Next month, we will examine another essential plotting technique: heatmaps.