Comparative Evaluation of Bipartite, Node-Link, and Matrix-Based Network Representations

Original link:

Network data is a common data type in daily life, such as social networks, paper citation networks, and biological networks. As shown in Figure 1, among the methods of network visualization, Node-link Diagram and Adjacency Matrix are the most commonly used. Another bipartite layout method (Bipartite Layout) was first used on bipartite graphs and later also extended to the visualization of ordinary static graphs. In bipartite graph layout, nodes are copied and placed on two parallel axes. , and connect them with edges. Each method has its strengths and weaknesses in different tasks on the web. This work [1] conducts a user study to evaluate these three network visualization methods and gives guidelines for their use.

Figure 1 Three network visualization methods

A total of 5 tasks were selected for this work, which can be divided into Overview Tasks and Detailed Tasks. Among them, the overview tasks include network type identification, cluster detection, and network density estimation; the detailed tasks include node in-out degree comparison and representation mapping (Representation Mapping). As shown in Fig. 2, for each task, this work generates network data of different densities and network types for user experiments to validate the hypotheses it proposes for each task.

Figure 2 Network visualization tasks and constructed datasets

Figure 3 shows the performance of different network visualization methods on each task. After hypothesis testing, the work draws the following conclusions. The node-link graph is inferior to adjacency matrix and bipartite graph visualization in the task of network type identification; bipartite graph visualization is worse than the other two in the task of cluster detection; in the task of network density estimation, there are significant differences in the effects of the three methods , the bipartite graph method is inferior to the other two; the adjacency matrix and the bipartite graph method are better than the node-link graph in the node in-exit comparison task; the adjacency matrix and the bipartite graph method have no significant difference in the representation mapping task .

Figure 3. Performance of different network visualization methods on different tasks (NL: node-link graph; AM: adjacency matrix; BP: bipartite graph)

Finally, this work gives the guidelines for the use of network visualization methods, as shown in Figure 4, different tasks are recommended to use different visualization methods. For example, for the task of comparing the overall structure of the network, it is recommended to use an adjacency matrix or a bipartite graph; for the task of clustering and community identification in the network, it is recommended to use a node-link graph; for tasks related to network density, it is recommended to use a node-link graph. Or an adjacency matrix; for tasks related to node attributes, adjacency matrices and bipartite graphs are recommended. In general, the adjacency matrix method appears to be more reliable in different tasks.

Figure 4 Summary of guidelines for using network visualization methods


[1] Moataz Abdelaal, Nathan D Schiele, Katrin Angerbauer, Kuno Kurzhals, Michael Sedlmair, Daniel Weiskopf. Comparative Evaluation of Bipartite, Node-Link, and Matrix-Based Network Representations. IEEE VIS 2022.

This article is reprinted from:
This site is for inclusion only, and the copyright belongs to the original author.