André Skupin and Sara Irina Fabrikant

s (Kohonen et al. 1999). Spring models are another popular category of dimensionality reduction techniques (Kamada and Kawai 1989, Skupin and Fabrikant 2003). Pathfinder network scaling (PFN) is a technique used for network visualization, with a preservation of the most salient links between input features. It is frequently applied to citation networks (Chen and Paul 2001). To illustrate this, we computed a PFN solution from the IJGIS citation data shown earlier. The result is a network structure consisting of links and nodes. When combined with a geometric layout of nodes derived from a spring model, the citation network can be visualized in GIS (Figure 9). Circle sizes represent the degree of centrality a paper has in this network, a measure commonly used in social network analysis (Wasserman and Faust 1999). Note how the centrality of the Takeyama/Couclelis paper derives from it being frequently cited (see Table 1), while the Wu/Webster paper establishes a central role because it cites a large number of IJGIS papers. [Figure 9 near here] Among spatial layout techniques, the treemap method has become especially popular in recent years. It takes a hierarchical tree structure as input and lays portions of it out in a given two-dimensional display space (Johnson and Shneiderman 1991). In the process, node attributes can also be visually encoded (Figure 10). For example, when visualizing the directory structure of a hard drive, file size could be encoded as the area size of rectangles. Another important category are graph layout algorithms, which attempt to untangle networks of nodes and links in such a manner that crossing lines are avoided as much as possible and network topology is preserved. [Figure 10 near here] Once dimensionality reduction or spatial layout methods have been applied, further 16 André Skupin and Sara Irina Fabrikant transformations are necessary to execute the visual design of a spatialization. Depending on the character of the base geometry, these transformations may include the derivation of feature labels, clustering of features, landscape interpolation, and others (Skupin 2002b, Skupin and Fabrikant 2003). When dealing with 2D geometry, much of this can be done in COTS GI Systems. Many aspects of these transformations remain to be investigated in future research, e.g., how scale changes can be implemented as semantic zoom operations (Figure 11). [Figure 11 near here] Spatialization geometry can also be linked to attributes that were not part of the input data set. For example, demographic change trajectories (Figures 1 and 2) could be linked – via symbolization or selection – to voting behavior or public policy decisions (Skupin and Hagelman 2005). 5 Usability and Cognitive Perspectives An extensive set of display techniques has been developed for spatialization, and the impressive array of visual forms documents the productivity of this young academic field (Chen 1999). However, few researchers have succeeded in providing empirical evidence to support claims that interactive visual representation tools indeed amplify people’s cognition (Ware 2000). Generally, non-expert viewers do not know how spatializations are created, and are not told through legends, or traditional map marginalia, how to interpret such aspects of spatialized displays as distance, regionalization and scale. Of the few existing experimental evaluations in information visualization, most evaluate specific depiction methods or types of software (Chen and Czerwinski 2000, Chen et al. 2000). While usability engineering approaches are good at testing users’ successes in extracting information from a particular