Data Mining - Mehmed Kantardzic [254]
Dimensional stacking is a recursive-visualization technique for displaying high-dimensional data. Each dimension is discretized into a small number of bins, and the display area is broken into a grid of subimages. The number of subimages is based on the number of bins associated with the two “outer” dimensions that are user-specified. The subimages are decomposed further based on the number of bins for two more dimensions. This decomposition process continues recursively until all dimensions have been assigned.
Some of the novel visual metaphors that combine data-visualization techniques are already built into advanced visualization tools, and they include:
1. Parabox. It combines boxes, parallel coordinates, and bubble plots for visualizing n-dimensional data. It handles both continuous and categorical data. The reason for combining box and parallel-coordinate plots involves their relative strengths. Box plots work well for showing distribution summaries. The strength of parallel coordinates is their ability to display high-dimensional outliers, individual cases with exceptional values. Details about this class of visualization techniques are given in Section 15.3.
2. Data Constellations. A component for visualizing large graphs with thousands of nodes and links. Two tables parametrize Data Constellations, one corresponding to nodes and another to links. Different layout algorithms dynamically position the nodes so that patterns emerge (a visual interpretation of outliers, clusters, etc.).
3. Data Sheet. A dynamic scrollable text visualization that bridges the gap between text and graphics. The user can adjust the zoom factor, progressively displaying smaller and smaller fonts, eventually switching to a one-pixel representation. This process is called smashing.
4. Time Table. a technique for showing thousands of time-stamped events.
5. Multiscape. A landscape visualization that encodes information using 3-D “skyscrapers” on a 2-D landscape.
An example of one of these novel visual representations is given in Figure 15.3, where a large graph is visualized using the Data Constellations technique with one possible graph-layout algorithm.
Figure 15.3. Data Constellations as a novel visual metaphor.
For most basic visualization techniques that endeavor to show each item in a data set, such as scatter plots or parallel coordinates, a massive number of items will overload the visualization, resulting in a clutter that both causes scalability problems and hinders the user’s understanding of its structure and contents. New visualization techniques have been proposed to overcome data overload problem, and to introduce abstractions that reduce the amount of items to display either in data space or in visual space. The approach is based on coupling aggregation in data space with a corresponding visual representation of the aggregation as a visual entity in the graphical space. This visual aggregate can convey additional information about the underlying contents, such as an average value, minima and maxima, or even its data distribution.
Drawing visual representations of abstractions performed in data space allows for creating simplified versions of visualization while still retaining the general overview. By dynamically changing the abstraction parameters, the user can also retrieve details-on-demand. There are several algorithms to perform data aggregations in a visualization process. For example, given a set of data items, hierarchical aggregation is based on iteratively building a tree of aggregates either bottom-up or top-down. Each aggregate item consists of one or more children that are either the original data items (leaves) or aggregate items (nodes). The root of the tree is an aggregate item