Related Work

The core part of our approach is a visual display of the language genealogy, which is a hierarchical data structure. According to Shneiderman's terminology [158] our data has a high fanout, that is, it potentially contains thousands of leaf elements, but the leaf-root distance is usually very short. Dierent approaches for plotting hierarchical data will be discussed in this section. For our purposes we need to compare multiple language features across languages within and across dierent hierarchical categories. To the best of our knowledge so far no other approaches have been published that pursue this as a main goal.

However, there are dierent approaches that plot relations among dierent nodes in a hierarchy, and approaches that combine geo-spatial information and hierarchical data.

Plotting hierarchical data

In the literature, several basic approaches for displaying hierarchical data can be found (1) Node-link tree diagrams, (2) Icicle Plots, (3) Treemaps, and (4)

Radial space-lling layouts.

(1) Node-link tree diagrams are the most intuitive and natural way of plotting hierarchies. In contrast to most other methods, node-link tree diagrams are not space-lling and many dierent layouts exist. Interesting extensions include the Hyperbolic Tree [102, 103], which lays out the hierarchy on a hyperbolic plane, and the three-dimensional animated Cone Tree [143]. The integrated change of focus interaction makes it a useful display for browsing large hier-archies. Node-link tree diagrams can also be plotted in a 3D space like in the case of the animated Cone Tree [143].

(2) Icicle plots [99] are space-lling rectangular versions of trees. The levels of the hierarchy are displayed as horizontal stripes from top to bottom and the elements of the hierarchy divide the stripes vertically into parts. Each hierarchy level requires the same amount of space as the leaf level.

(3) Both the concepts of nested and non-nested Treemaps have rst been pub-lished by Johnson and Shneiderman [87,157] and since then have become very popular and were extended in dierent ways for dierent purposes. A brief overview of the history of Treemaps by Shneiderman and Plaisant can be found online⁴. Interesting extensions include Ordered and Quantum Treemaps [18], Voronoi Treemaps [15], and Generalized Treemaps [173]. Treemaps grant al-most the whole display space to the leaf nodes making internal nodes in non-nested Treemaps only visible as space separators, i.e.,in Treemaps the hierarchy is conveyed through containment. They show their strengths when the focus of analysis lies on the leaf nodes and especially when these leaf nodes have dierent sizes that are important to explore. The shapes of leaf nodes may dier considerably, even if they are all granted the same space.

(4) Radial space-lling layouts like the Information Slices [8] and later the Sun-burst display [161] have the advantage that they do not grant as much space to the inner nodes as Icicle Plots, but still have a space-lling representation for them. Typically, the amount of nodes in a hierarchy increases with increasing distance to the root. In the Sunburst visualization the amount of display space available at each level of the hierarchy increases analogously.

Other visualizations for hierarchical data can be found but have not become as popular as the aforementioned ones. Examples include the Cheops system

4http://www.cs.umd.edu/hcil/treemap-history/ last revised on March 6th, 2013

[17] that emphasizes on browsing and exploration of complex hierarchies but not on the analysis.

Hierarchical and relational data

A visualization that integrates both hierarchical and relational data are Arc-Trees [128], a combination of a Treemap with an arc diagram. The Treemap grows only in horizontal direction and linking arcs connect two nodes of the hierarchy if they are related. Another technique with a similar purpose is Holten's Hierarchical Edge Bundles [79], which can be combined with dierent hierarchical visualizations. Elements in the hierarchy are connected with col-ored bundled links if they are related. Both ArcTrees and Hierarchical Edge Bundles, however, require the link space to be rather sparsely populated and do not scale for fully connected graphs. In addition, while it can be conveyed that two items are related, dierent types of relations are hard to express and the links are not suitable for performing feature comparison across elements in the hierarchy.

A further possibility of combining hierarchical data with relational clues is to provide a matrix display that shows relations in the cells. Either the axes elements of the matrix are leaf nodes of a hierarchical node-link tree structure as in the Matrix Browser [189] or the hierarchy is conveyed through a Treemap-like recursive subdivision of the matrix as in [169]. Fully connected graphs are not a problem for these latter approaches, but overall only a limited number of nodes can be displayed; otherwise the matrix will grow too big. For all of the mentioned techniques, there is no intuitive way to use the visualization for feature comparison.

Hierarchical and geo-spatial data

An approach that combines hierarchical and geo-spatial data are Flow Maps [24, 137]. Flow Maps lay a tree structure over a map in order to indicate geo-spatial movements (ows). The tree structure is the result of a hierarchical clustering of locations. Thus, the hierarchy directly depends on the geo-locations and is of a binary nature. In our scenario, in contrast, the hierarchy is predened and not directly related to geo-spatial distributions. In addition, the authors state that good ow maps contain a moderate number of nodes (less than

100) [137], which is not the case for our data. Recently, an improved layout based on spiral trees has been introduced [24]. A dierent approach to combine geographic and hierarchical information into one visual display is to consider spatial ordering when creating space-lling rectangular layouts like Treemaps.

One option is to take longitude and latitude values into account when splitting the Treemap rectangles as in Mansmann's Geographic HistoMap Layout [117].

Wood and Dykes [182] follow the same fundamental idea with their Spatially Ordered Treemaps. Later Slingsby et al. [159] suggest a further version of geographically ordered space-lling rectangular layouts. All of the mentioned approaches share the property that either the upper levels of the hierarchy are geospatial, e.g., areas and subareas, or that in each leaf node geo-spatial distributions have to be displayed. The latter option causes diculties when having many geo-spatial locations and limits the possibilities for conveying feature information in leaves, because the most important visual variables are already used to convey the geo-spatial dimensions.