Academic journal article Cartography and Geographic Information Science

The Distance-Similarity Metaphor in Network-Display Spatializations

Academic journal article Cartography and Geographic Information Science

The Distance-Similarity Metaphor in Network-Display Spatializations

Article excerpt

Introduction

Spatialization is the process of generating an information display of non-spatial data. Information spatialization is inspired by the intuition that spatial or graphical displays (e.g., maps, charts, photographs, diagrams) can help to amplify cognition (Tversky 2000), whether or not viewers are familiar with the construction details of the viewed display (Card et al. 1999). Generally, non-expert viewers do not know how spatializations are created and are not told through legends or other traditional map marginalia how to interpret the displays.

Spatializations typically rely on dimension reduction techniques (e.g., multidimensional scaling, self-organizing maps) and layout algorithms (e.g., energy minimization/force-feedback models) to project relatedness (e.g., similarity) in non-spatial data content onto distance, such that semantically similar documents are placed closer to one another than are less similar ones in an information space (Borner et al. 2003). We have coined this widely applied design principle the "distance-similarity metaphor" (Montello et al. 2003). In information visualizations (Chen 1999), especially those depicting knowledge-domains (Chen 2003), node-link displays are popular graphical devices for expressing the distance-similarity metaphor. A good example is the TouchGraph Google Browser (1), which provides a node-link display as a graphical user interface to Google's "what's related" search option, used to find web pages similar to a queried page.

Figure 1 depicts a network of web pages related to the highlighted web page labeled "Saraland" in the center of the display (the web page belongs to the first author). The placement of the nodes is achieved with a force-feedback graph-layout algorithm (e.g., Kamada and Kawai 1989). Nodes are re-arranged in a spring-like fashion, so that connected web sites not only get linked, but more strongly connected web sites contract towards each other in the spatialization. The resulting configuration is modified aesthetically, including reducing edge crossings to a minimum and evening edge lengths to achieve a balanced and visually pleasing layout.

[FIGURE 1 OMITTED]

Especially popular as a way to depict semantic relatedness has been the Pathfinder Network Scaling (PFNET) approach, derived from author co-citation analyses or word co-occurrences in text documents. Dominant relationships in proximity data are represented as a 2D or a 3D graph in PFNETs, and when depicted in 2D, the PFNET looks like a network map, similar to the one depicted in Figure 1. PFNET computes a pattern of links based on pair-wise assessment of proximity information associated with nodes in the graph. Although all the nodes are included in the PFNET, the network is typically computed with a minimum of links, such that only the strongest proximity relationships between nodes are depicted, as shown in Figure 2.

Figure 2 depicts a subset of a larger PFNET of Reuters news articles. The whole network contains 504 documents of randomly selected news stories, collected during February 9-10, 2000 (Fabrikant 2001). The nodes represent individual news articles and the links represent semantic relationships between the articles based on an analysis of word co-occurrence computed by the latent semantic analysis techniques (LSI) (Deerwester et al. 1990). The labels were extracted automatically from the article to indicate the main themes of the documents (2). In essence, a PFNET graph is akin to a minimum spanning tree. Which nodes get linked is determined by a chosen direct-distance metric, but unlike the minimum spanning tree, the triangle inequality requirement for metric spaces is relaxed (Schvaneveldt 1990).

[FIGURE 2 OMITTED]

Background

Point-display spatializations (e.g., multidimensional scaling plots) depict documents as unlinked points in space, semantic relatedness being reflected in the proximity relationships of the points (Montello et al. …

Search by... Author
Show... All Results Primary Sources Peer-reviewed

Oops!

An unknown error has occurred. Please click the button below to reload the page. If the problem persists, please try again in a little while.